Associations Between Blood Leukocyte DNA Methylation and Sustained Attention in Mid-To-Late Childhood

Abstract

Aims: To identify associations between DNA methylation (DNAm) across the epigenome and symptoms related to attention-deficit/hyperactivity disorder in a population of Hispanic children. Materials & methods: Among 517 participants in the ELEMENT study aged 9–18 years, we conducted an epigenome-wide association study examining associations between blood leukocyte DNAm and performance on the Conners’ continuous performance test (CPT3). Results: DNAm at loci in or near ZNF814, ELF4 and OR6K6 and functional enrichment for gene pathways pertaining to ferroptosis, inflammation, immune response and neurotransmission were significantly related to CPT3 scores. Conclusion: DNAm was associated with CPT3 performance. Further analysis is warranted to understand how these genes and enriched pathways contribute to attention-deficit/hyperactivity disorder.

Keywords::

• adolescent health
• child health
• DNA methylation
• epigenetic epidemiology
• epigenetics
• neurobehavior
• neurodevelopment

Attention-deficit/hyperactivity disorder (ADHD) is one of the most common neurodevelopmental disorders, affecting roughly 5.3% of children globally. Two thirds of people diagnosed with ADHD in childhood experience symptoms into adulthood [1]. The disorder can impact multiple parts of the brain and is associated with increased risk of cardiovascular diseases, accidental self-injury, substance abuse, depression, anxiety, eating disorders, suicides, incarcerations and obesity. ADHD is characterized by symptoms of impulsivity, hyperactivity and inattentiveness. Because its effects on executive functioning often result in poor self-regulation and decision-making, the affected individual may engage in risky behavior that predisposes them to adverse health effects. People with childhood ADHD are twice as likely to die prematurely than those without ADHD, and if ADHD persists into adulthood, they have a reduced estimated life expectancy of about 13 years with a 9.5-year reduction in healthy estimated life expectancy [1,2]. Despite these findings, ADHD remains underacknowledged as a public health concern, and continues to increase in prevalence [3]. To address this issue, we need to not only conceptualize ADHD as a risk factor for other health conditions, but also understand the risk factors for the disorder itself.

Epigenetic mechanisms, including DNA methylation (DNAm), are potential contributors to ADHD susceptibility and etiology. Epigenetic modifications may alter gene expression but do not change the underlying genetic code. They are fairly stable and heritable, at least across cell divisions (mitosis), but are still modifiable, making them an excellent target for potential intervention. In the field of epigenetics, DNAm is the most commonly studied process in humans. It occurs when a methyl group is added to a cytosine base located 5 prime of a guanine base referred to as a CpG site. A body of literature exists relating DNAm to neuro-developmental and -behavioral outcomes [4–8]. DNAm at CpG sites, especially in promoter regions of genes, is predominantly associated with gene silencing. DNAm is known to help regulate brain development and is hypothesized to play a role in ADHD [4,5]. In a recent study conducted by Li et al., researchers used a continuous performance test (CPT), a commonly used method to assess ADHD-like behaviors in children, to explore the relationship between DNAm and attention-related symptoms in a sample of Taiwanese children [5]. They found significant associations between higher DNAm levels (hypermethylation) at the LIME1 gene and poorer CPT test performance. They also documented a significant relationship between hypomethylation at the SPTBN2 gene and worse CPT scores [5]. Both LIME1 and SPTBN2 are implicated in neurological disorders. LIME1 plays a major role in inflammatory pathways like MAPK signaling and has been linked to cerebral palsy pathophysiology [9]. SPTBN2, which helps regulate the glutamate signaling pathway through stabilization of the glutamate transporter EAAT4, is associated with spinocerebellar type 5 in humans [10,11]. The Li et al. study was conducted using candidate genes whose DNAm levels were identified as ADHD relevant genes through a pilot epigenome-wide association study (EWAS) (data from 12 ADHD cases and 9 controls). Selected candidate sites that were associated with ADHD in the EWAS and related to any neurological process were subjected to further analysis via pyrosequencing using a sample of 198 children (126 ADHD patients and 72 controls) [5]. In a larger EWAS by Mooney et al. (391 cases and 213 neurotypical controls) DNAm was found to be associated with ADHD diagnosis and polygenic risk score [6]. This study was conducted in a US population of children aged 7–12 years. Other EWASs have been conducted, exploring the relationship between DNAm and ADHD-like symptoms in children [7,8]. Collectively these studies suggest that DNAm may be part of the etiology of ADHD. However, past studies are limited to primarily Asian and White populations. Larger studies in more ethnically diverse populations are needed to validate past findings and identify differences across cohorts of varying demographics. One of such populations was used in the present study employing data from similarly aged participants of the Mexican ELEMENT cohort.

Childhood and adolescence are important neurodevelopmental periods. Development of executive functioning skills, which is closely related to maturation of the frontal lobe in the brain, does not generally begin to emerge until about 6 years of age and continues to develop through adolescence [12]. Not only do epigenetic mechanisms help regulate brain development in childhood and adolescence, but they are hypothesized to play a role in ADHD [4,5]. Only a handful of EWASs currently explore the relationship between DNAm and executive function at these highly relevant developmental periods [5–8]. Identifying genes that are differentially methylated in relation to ADHD may help us identify potential biomarkers and enable us to gain insight about mechanistic pathways. To address these gaps, we assessed the associations between blood leukocyte DNAm using an epigenome-wide screen and CPT3 performance in children and adolescents of Latino origin from the Mexico City-based ELEMENT cohorts.

Materials & methods

Cohort

A total of 1012 pregnant mothers in Mexico City were recruited (1994–2004) from prenatal clinics from the Mexican Social Security Institute, serving low- to middle-income populations. Mother/child dyads were followed through present day. Data were collected at various timepoints, including in 2015, when a subset of 550 children (between 9 and 18 years of age) participated in a follow-up visit [13]. This study focuses on 526 (51% female and 49% male) participants at this timepoint, for whom epigenome-wide DNAm was quantified in blood leukocytes. At this timepoint, most participants also completed the Conners’ continuous performance test, third edition (CPT3). Approximately 517 participants had both CPT3 performance data and DNAm data collected and are included in this analysis. Research protocols were approved by the institutional review boards of the Mexico National Institute of Public Health and the University of Michigan. Parental informed consent and child assent was obtained for all participants in this study.

Epigenetic analysis

All participants attending the adolescent study visit and providing a blood sample were selected for epigenome-wide DNAm analysis as previously described [14]. To generate methylation data, we extracted DNA from blood leukocytes via Qiagen Flexigene kits (Hilden, Germany), performed bisulfite conversion on the samples and subsequently analyzed them on an Illumina Infinium Methylation EPIC BeadChip (CA, USA) (referred to as the EPIC array). This probe-based assay quantifies methylation at more than 850,000 CpG sites across the human genome [15]. While a fraction of those comprising the entire genome (∼28 million CpG sites), it is representative of a vast distribution of regions within and outside of genes, including promoters, enhancers and intergenic regions, providing coverage at loci within every known gene [16].

Bisulfite-converted DNA samples from the ELEMENT participants were randomized across chips and positions and hybridized to BeadChips. Signals were read, processed and subjected to rigorous quality controls at the University of Michigan Advanced Genomics Core. To analyze this data, we followed a standardized pipeline for this technology, previously described [14]. In brief, we read raw image data files into R using the minfi package and corrected for background noise and dye bias using RELIC [17,18]. To account for unwanted technical variability, estimated from control probes on each chip, quantile normalization was performed [18,19]. We excluded probes with poor detection above background (in less than 5% of samples), probable crossreactivity, and/or polymorphisms in the CpG or single base extension sites [20]. Additional quality control measures included comparison of documented gender to that estimated from X and Y chromosomal DNAm profiles, low intensity across all probes and/or >5% probes failing. An initial quality control check showed that all 526 adolescent samples undergoing this analysis passed quality control. As heterogenous populations of cells comprise whole blood (the sources from which the DNA was extracted), cell type proportions were estimated based on data from tissue-specific differentially methylated regions (DMRs) included on the chips [21]. Surrogate variable analysis was performed using the intensity values from the non-negative control probes to create variables representing technical variation influencing the DNAm data [22]. The final data used in downstream statistical analysis reflect average β-values for the proportion of total methylated cytosine at each of the 762,657 CpG sites included in this analysis.

Cognitive performance test

At the adolescent study visit, CPT3 was conducted by trained research staff. For this computerized test, children are presented with a series of stimuli to which they respond by either hitting a button or not. From the response data scores, variables are generated to reflect four different aspects of attention: inattentiveness, impulsivity/hyperactivity, sustained attention and vigilance [23,24], as shown in Table 1. Four representative variables for each aspect of attention were selected as outcomes based on ease of interpretation and prevalence of use in the literature. Response style, generated by the CPT3 test to account for individualistic style of responding, was considered a covariate in this study [23–25].

Table 1. Descriptions of Conners’ continuous performance test (third edition) variables used in this study.

Variable	Definition	Aspects of attention	Interpretations	Notes on data transformation and interpretation
Omissions	Percent of missed targets	Inattentiveness	Higher T scores reflect greater rate of omission errors	This variable was log transformed to meet assumptions for linear models
Perseverations	Percent of random or anticipatory responses (HRT: <100 ms, the minimum processing time for informed response)	Impulsivity	Higher T scores reflect greater rate of perseveration errors	May be related to impulsivity or very liberal response style, so response styles were controlled for in models. This variable was log transformed to meet assumption for linear models
HRT block change	Change in response speed across the six blocks	Sustained attention^†	Higher T scores reflect slower response speeds in the later blocks, which reflects a reduction in information processing efficiency over time (loss of sustained attention)	Not log transformed
HRT ISI change	Change in response speeds at various interstimulus intervals (ISI: 1 s, 2 s and 4 s)	Vigilance^†	Higher T score signify greater slowing at longer ISIs, which reflects a reduction in information processing efficiency (loss of vigilance)	Not log transformed
Response style^‡	Signal detection variable assessing an individual’s natural tendencies to respond to stimuli based on a speed vs accuracy tradeoff		Higher T scores indicate conservative response styles (emphasis on accuracy over speed) and lower scores indicate liberal response styles (emphasis on speed over accuracy)	Response style can influence one’s scores for other variables (i.e., HRT and commissions errors). Adjusted for in models

† Some scientists consider vigilance and sustained attention to be synonymous, but Conners’ continuous performance test, third edition (CPT3) recently acknowledged them as distinct items with vigilance describing one’s ability to remain alert even in situations with less stimulation (i.e., with a longer ISI) [23–25].

‡ This variable was also generated by the CPT3, but was included as a covariate rather than an outcome variable in each model

HRT: Hit reaction time; ISI: Interstimulus interval.

Covariates

Additional variables, including child gender, age, CPT3 response style and socioeconomic status (SES), were considered as potential confounders in the relationship between DNAm and CPT3 outcomes on the basis of their biological relevance and/or correlation (Spearman) with the outcomes (Supplementary Table 1). SES was quantified using The Mexican Association of Marketing Research and Public Opinion Agencies (AMAI) scaling system, specifically designed for Mexican populations, as a metric [26,27]. In subsequent analyses, principal component 4, representing batch effects from the EPIC data, and cell type proportions in whole blood (CD8⁺ T cells, B cells and granulocytes) were added to the models. These variables were selected as described below. Child Tanner staging, a measure of pubertal status determined by a trained physician at the study visit, was also considered but not included in the models, as this variable was not associated with the outcome variables and its inclusion in the model did not improve fit.

Statistical analysis

All statistical analyses were performed in R version 4.0.3. Linear regression models were employed to identify CpG sites with methylation associated with CPT3 performance (using outcome variables omission errors, perseverations, hit reaction time interstimulus interval change and hit reaction time block change in independent analyses). Main models were adjusted for age, gender, SES (AMAI) and response style. For sensitivity analyses, an additional model was run for each of the four outcomes, controlling for age, gender, response style, principal component 4 representing batch effects from the EPIC data and cell type proportions (CD8⁺ T cells, B cells and granulocytes). These variables were selected from a singular value decomposition analysis, which assessed the correlation between each potential covariate and principal components representing DNAm across all probes (Supplementary Figure 1). These additional variables explain a large portion of the DNAm data variance but are not expected to have a substantial influence on the outcome, which is why precedence was placed on biological relevance and correlation with CPT3 outcomes in selection of covariates for the main model analyses. Gender-stratified models were also conducted, accounting for age, SES and CPT3 response style. CpG sites were considered statistically significantly associated with outcomes using Benjamini–Hochberg false discovery rate (q-values < 0.2) [28]. This significance threshold is only moderately stringent, so results that are statistically significant at the more conventional q < 0.05 are also highlighted. The 0.2 cutoff was selected because it has been used in other epigenome-wide studies of neurologically related outcomes [29–33]. The missMethyl R package, accounting for unequal representation of CpG sites within genes on the EPIC array, was used to identify functional pathways for differentially methylated genes enriched among the loci with the smallest raw p-values (top 1000) for each outcome using both Kyoto Encyclopedia of Genes and Genomes (KEGG) and GO (Gene Ontology) pathways.

Gene expression data

To quantify the relative expression of genes across the entire genome, RNA sequencing was conducted for a subset of participants with RNA isolated from blood (n = 72). Expression data were obtained using paired-end 50-cycle sequencing on an Illumina HiSeq 4000 (CA, USA) at the University of Michigan Advanced Genomics Core, and was processed and normalized as previously described [14]. For the 69 participants for whom both gene expression and DNAm data exist, correlations between DNAm at each of these probes and expression of the gene annotated to these loci were calculated. In addition, correlations between DNAm at significant differentially methylated sites (associated with any of the four outcomes) and genome-wide expression was identified via Spearman correlation using a Bonferroni corrected critical value of 3.138E-06 (0.05/15932).

Blood–brain DNAm comparison

Employing a publicly available tool, generated by Hannon et al., comparisons were made between DNAm in blood and DNAm in various brain regions for each of the statistically significant probes identified to be differentially methylated by CPT3 performance in the present study [34]. This blood–brain DNAm comparison tool was created using DNAm data generated via the Illumina 450K array (previous version of the EPIC array) from matched blood and brain samples collected postmortem from adults. Out of the five significant loci identified in our study, three were available in the database and evaluated.

Results

Of the 517 adolescent participants for whom epigenetic data and CPT3 results were available, 51% were female and 49% were male. CPT3 scores exhibited a broad range of values and were comparable to cohorts of similarly aged children across the globe (Table 2). After correcting for multiple comparisons via the Benjamini–Hochberg method using a significance threshold of 0.2, four significant probes were identified that were associated with interstimulus interval change: cg02280912, cg12603272, cg22510337 and cg22601108 (Table 3 & Supplementary Table 2). One of these probes (cg22510337 in gene ZNF814) was also significantly associated with block change in both the primary model and sensitivity analysis. DNAm at this probe, along with cg23825347, was significantly associated with perseverations, but only in the sensitivity model (Table 3, Supplementary Tables 2 & 3). p-values from the models for the outcome quantifying inattention (omissions error) were likely inflated (λ = 1.99). For all outcomes other than omissions error, a general overview of the number of statistically significant findings for main models and sensitivity models is outlined in Supplementary Table 2. All statistically significant results from the main models and sensitivity models are presented in Table 3 & Supplementary Table 3, respectively. When models were stratified by gender, we identified only two probes (cg13097337 and cg23024411) whose methylation status was significantly associated with CPT3 score, specifically the block change outcome. Both were significant in females (cg13097337, β: -3.014; standard error [SE]: 0.582; q = 0.17; cg23024411, β: 1.380; SE: 0.266; q = 0.17), but not in males (cg13097337, β: -0.070; SE: 0.24; q = 1.00; cg23024411, β: -0.209; SE: 0.432; q = 1.00) (Supplementary Table 4).

Table 2. Descriptive statistics for adolescent ELEMENT participants included in this analysis.

Variable	n	Median	Interquartile range	Mean	Standard deviation	Range
Age (years)	517	14.4	12.8–16.7	14.59	2.12	9.8–18.1
Response style	517	52	47–58	52.45	9.27	27–90
Omissions error	517	47	44–53	50.57	10.48	41–90
Perseveration	517	46	45–52	51.17	11.35	44–90
Block change	516	51	45.75–57	51.5	9.84	13–90
Interstimulus interval change	517	50	44–56	51.03	10.23	27–90
Gender		n (%)
Male		243 (47)
Female		274 (53)
AMAI
A/B		3 (0.58)
C+		35 (6.77)
C		94 (18.18)
C-		126 (24.37)
D+		125 (24.18)
D		132 (25.53)
E		2 (0.39)

For the AMAI variable, categories are listed sequentially with A/B representing the highest socioeconomic status level (A/B = 7) and category E reflecting the lowest socioeconomic status level (E = 1).

AMAI: The Mexican Association of Marketing Research and Public Opinion Agencies scaling system.

Table 3. Loci with DNA methylation statistically significantly associated with outcomes of Conners’ continuous performance test, third edition.

Probe ID	Outcome	Estimate	SE	p-value	Q-value	Gene name	Chromosome	Position	Relation to island	University of California Santa Cruz RefGene annotation
cg02280912	ISI change	0.719	0.142	5.49E-07	0.122	ZNF814	chr19	58400095	Island	Body
cg12603272	ISI change	0.873	0.173	6.38E-07	0.122	ELF4	chrX	129221241	Open sea	5′UTR
cg22510337	ISI change	0.909	0.18	5.80E-07	0.122	ZNF814	chr19	58399967	Island	Body
cg22601108	ISI change	1.815	0.358	5.54E-07	0.122	OR6K6	chr1	158723939	Open sea	TSS1500
cg22510337	Block change	0.971	0.169	1.55E-08	0.012	ZNF814	chr19	58399967	Island	Body

ISI: interstimulus interval; SE: Standard error.

For full participant models, we tested for enrichment in functional gene pathways from the KEGG and GO databases. For the interstimulus interval (ISI) change outcome, the top KEGG pathways and GO terms (up to top 25 lowest p-values where p < 0.05) are presented in Tables 4 & 5, respectively. The results for the remaining outcomes are included in the supplemental materials (Supplementary Tables 5–10). Across all outcomes, enrichment for pathways related to neurotransmission, immune response and inflammation was observed. Specific KEGG pathways were enriched for multiple outcomes. For instance, the glutamatergic synapse term was enriched in the block change and ISI change EWASs (Supplementary Table 5 & Table 4), ferroptosis was enriched for perseveration and ISI change EWASs (Supplementary Table 7 & Table 4), and MAPK signaling pathway was enriched for block change and perseveration EWASs (Supplementary Tables 5 & 7). Notably, after employing a q-value cutoff of q < 0.2, few pathways were significantly enriched for any outcome. The KEGG pathway ferroptosis was enriched among results from the ISI change analysis (q-value = 0.048).

Table 4. Top Kyoto Encyclopedia of Genes and Genomes terms enriched among results from the interstimulus interval change model.

KEGG term	Description	n^†	No. differentially methylated genes	p-value	FDR q-value	Significant in sensitivity model, too?
path:hsa04216	Ferroptosis	41	8	1.00E-04	0.048^‡	Y
path:hsa00061	Fatty acid biosynthesis	17	5	0.001	0.193	N
path:hsa03040	Spliceosome	130	11	0.002	0.193	Y
path:hsa04658	Th1 and Th2 cell differentiation	90	10	0.005	0.467	N
path:hsa04723	Retrograde endocannabinoid signaling	141	13	0.009	0.554	N
path:hsa04114	Oocyte meiosis	121	11	0.01	0.554	N
path:hsa00071	Fatty acid degradation	43	5	0.011	0.554	N
path:hsa04914	Progesterone-mediated oocyte maturation	92	9	0.02	0.799	N
path:hsa05142	Chagas disease	102	9	0.021	0.799	N
path:hsa04936	Alcoholic liver disease	139	10	0.03	0.802	N
path:hsa01212	Fatty acid metabolism	56	6	0.03	0.802	N
path:hsa04912	Gonadotropin hormone-releasing hormone signaling pathway	93	9	0.035	0.802	N
path:hsa04933	AGE–RAGE signaling pathway in diabetic complications	100	9	0.038	0.802	Y
path:hsa04218	Cellular senescence	155	12	0.041	0.802	N
path:hsa05150	Staphylococcus aureus infection	86	5	0.042	0.802	N
path:hsa04520	Adherens junction	71	8	0.044	0.802	N
path:hsa04724	Glutamatergic synapse	114	11	0.049	0.802	N

p-values were ranked from lowest to highest. Up to the 25 terms with the lowest p-values (must be p < 0.05) were included in this table.

† Number of genes included in the GO or KEGG term.

‡ Statistically significant q-value with a significance threshold of q < 0.05.

FDR: False discovery rate; N: No; Y: Yes.

Table 5. Top Gene Ontology terms enriched among results from the interstimulus interval change model.

Description	Term identifier	Ontology	n^†	No. differentially methylated genes	p-value	FDR q-value	Significant in sensitivity model?
Negative regulation of leukocyte degranulation	GO:0043301	BP	12	4	4.32E-04	1	Y
Cellular response to histamine	GO:0071420	BP	8	4	5.45E-04	1	N
Recycling endosome	GO:0055037	CC	187	18	8.18E-04	1	N
Negative regulation of exocytosis	GO:0045920	BP	32	6	1.08E-03	1	Y
Ethanol catabolic process	GO:0006068	BP	11	3	1.34E-03	1	N
Response to histamine	GO:0034776	BP	11	4	1.53E-03	1	N
Primary alcohol catabolic process	GO:0034310	BP	14	3	2.61E-03	1	N
CD4^+, CD25^+, α/β-regulatory T-cell lineage commitment	GO:0002362	BP	2	2	3.07E-03	1	Y
Negative regulation of myeloid leukocyte mediated immunity	GO:0002887	BP	10	3	3.48E-03	1	Y
Nuclear speck	GO:0016607	CC	393	27	3.78E-03	1	N
Regulation of atrial cardiac muscle cell membrane repolarization	GO:0060372	BP	8	3	4.12E-03	1	N
Synapse pruning	GO:0098883	BP	11	3	4.72E-03	1	N
Negative regulation of leukocyte mediated immunity	GO:0002704	BP	52	6	5.16E-03	1	N
Cell junction disassembly	GO:0150146	BP	19	4	5.26E-03	1	N
Positive regulation of protein lipidation	GO:1903061	BP	4	2	5.40E-03	1	N
Regulation of vascular permeability	GO:0043114	BP	45	7	5.58E-03	1	N
SH3 domain binding	GO:0017124	MF	130	14	5.75E-03	1	N
Double-strand break repair involved in meiotic recombination	GO:1990918	BP	5	2	5.99E-03	1	N
Positive regulation of reactive oxygen species metabolic process	GO:2000379	BP	104	10	6.46E-03	1	N
Negative regulation of cellular extravasation	GO:0002692	BP	9	3	6.47E-03	1	N
Negative regulation of neutrophil degranulation	GO:0043314	BP	2	2	6.87E-03	1	Y
CD8^+, α/β-T-cell lineage commitment	GO:0043375	BP	2	2	6.88E-03	1	Y
Regulation of reactive oxygen species metabolic process	GO:2000377	BP	197	15	7.25E-03	1	Y
Positive regulation of mitochondrial ATP synthesis coupled electron transport	GO:1905448	BP	5	2	7.30E-03	1	N
Negative regulation of vascular permeability	GO:0043116	BP	16	4	7.39E-03	1	N

p-values were ranked from lowest to highest and the 25 terms with the lowest p-values that also had p < 0.05 were included in this table.

† Number of genes included in the GO or KEGG term.

BP: Biological process; CC: Cellular component; FDR: False discovery rate; GO: Gene Ontology; MF: Molecular function; N: No; Y: Yes.

For a subset of ELEMENT participants, transcriptomics gene expression data were also obtained from whole blood [35]. Correlations between DNAm at loci significantly associated with outcomes and gene expression of their annotated genes were generally not statistically significant (Table 6). When examining Spearman correlations between DNAm at these significant probes and expression across all genes, significant correlations (Bonferroni corrected p < 3.138E-06) were only identified between DNAm at one probe (cg12603272) and gene expression of 19 genes (Supplementary Table 11). Because this probe was located on the X chromosome, the correlation analysis was repeated in males and females separately. When stratified by gender, no correlations were statistically significant.

Table 6. Spearman correlation between DNA methylation at a given probe and expression of its nearest gene.

Gene\Probe ID	cg22510337	cg02280912	cg12603272	cg22601108
ZNF814	-0.21 (0.08)^†	-0.13 (0.3)	NA	NA
ELF4	NA	NA	0.08 (0.52)	NA
OR6K6	NA	NA	NA	-0.15 (0.21)

r(p), where r = correlation coefficient and p = p-value.

† p < 0.10.

NA: Not applicable (did not test).

Using the blood–brain DNAm comparison tool, we identified significant correlations between DNAm from matched blood and brain samples that are part of the database at some of the loci associated with CPT3 outcomes in our study. DNAm in blood at both of the CpG sites for ZNF814 that we found to be significantly associated with aspects of attention were correlated with DNAm from different brain regions: whole blood DNAm at cg02280912 and cg22510337 were correlated with DNAm in the superior temporal gyrus (r = 0.355; p = 0.002) and the entorhinal cortex (r = 0.348; p = 0.003), respectively. Lastly, for probe cg22601108, located in the OR6K6 gene, DNAm in whole blood and the prefrontal cortex were correlated (r = 0.209; p = 0.073) [34].

Discussion

The etiology of ADHD continues to be poorly understood, but DNAm in peripheral tissues of children has been associated with this disorder. In a cohort of Mexican children aged 9–18 years, we identified relationships between peripheral blood DNAm at specific loci and performance on the CPT3 test, which quantifies measures of inattention, impulsivity, sustained attention and vigilance. At a significance threshold of q = 0.2, we identified five significant probes, including loci annotated to the genes ZNF814, ELF4 and OR6K6. The direction and magnitude of effect sizes for these loci were comparable in the primary model and in a sensitivity analysis adjusting for batch and cell type proportions, providing further confidence in our results.

One gene of particular interest is ZNF814, as two loci annotated to this gene were significantly associated with outcomes (probe IDs cg22510337 and cg02280912, both loci located near each other in the first intron of the gene in a known CpG island and active regulatory region; Supplementary Figure 2) [36]. One of these, cg22510337, was associated with three out of four CPT3 outcomes: block change, ISI change and perseverations. ZNF814 is a protein-coding gene encoding ZNF814, which is believed to help facilitate DNA-binding transcription activator activity, as well as RNA polymerase II-specific and RNA polymerase II cis-regulatory region sequence-specific DNA binding activity. It is also suspected to be involved in regulating transcription via RNA polymerase II. Genetic variations in ZNF814 have been documented in patients with familial hemangioblastomas – rare, slow-growing tumors of the CNS [37]. Polymorphisms in several other zinc finger proteins in this family (C₂H₂-type zinc finger proteins [C₂H₂-ZNFs]) have been associated with neurodevelopmental disorders, including bipolar disorder, schizophrenia and autism spectrum disorders [38,39]. To our knowledge, zinc finger proteins have not been shown to be associated with ADHD specifically, but C₂H₂-ZNFs are known to be highly expressed in developing human brain tissue and help regulate early CNS gene expression patterning [40,41]. In one study, a SNP in the ZNF804A gene was found to be associated with executive control of attention in healthy volunteers, but in subsequent studies within ADHD participants researchers were unable to identify significant associations between this zinc finger protein variant and ADHD [38,42,43]. This discrepancy could be due to the fact that one study broke down attention into symptom profiles while the other utilized a broader clinical diagnosis of ADHD. Differences in demographic characteristics between the study populations, for example, age, race/ethnicity, SES and disease state, may also have played a role. Regardless, evidence from this and other published literature supports the idea that C₂H₂-ZNFs, which are the largest family of transcription factors in humans, may be important targets for the pathologic processes associated with mood and neurodevelopment disorders, including ADHD.

Whole blood is often used as a proxy for the target tissue of interest in epigenetic studies. However, DNAm can vary by tissue and even cell type, so assessing whole blood DNAm may not reflect the same findings we would see in brain regions and cells. We accessed publicly available resources comparing DNAm from matched blood and brain samples to infer whether associated genes have correlated or divergent DNAm between the surrogate and target tissue of interest. This blood–brain DNAm comparison tool generated by Hannon et al. is a unique resource; however, it must be noted that brain tissue was collected postmortem and may not be representative of DNAm profiles in healthy brain tissue. Regardless, utilization of this tool for probes we found to be significantly associated with aspects of attention generated some interesting findings. At the ZNF814 probe cg02280912, for example, blood DNAm was correlated with DNAm in the superior temporal gyrus [34]. The superior temporal gyrus is a section of the brain that is involved in auditory processing, language and social cognition [44]. Abnormalities in the superior temporal gyrus (i.e., increased radial diffusivity) have been documented in children with ADHD [45,46]. In another example, whole blood DNAm at cg22510337, a probe that is also annotated to ZNF814, and DNAm at the same probe in the entorhinal cortex were associated with each other [34]. The entorhinal cortex plays a major role in regulation of the neurocircuitry of the hippocampus, the part of the brain responsible for learning and memory; both structures are implicated in the etiology of neuropsychiatric disorders such as ADHD [47,48]. Another important structure implicated in the etiology of ADHD is the prefrontal cortex [48–50]. We found that blood and prefrontal cortex DNAm were correlated at cg22601108 (loci in the OR6K6 gene) (r = 0.209; p = 0.073) [34]. While blood leukocytes and brain cells are innately designed to perform different functions and are known to have different epigenetic regulation at many genes, the correlations identified between blood and brain tissues may represent a common set of genes with similar epigenetic patterns in both tissues. These genes may serve as better biomarkers of ADHD-like symptoms than others in blood. Another potential explanation of the association between blood DNAm and ADHD symptoms comes from emerging evidence suggesting that various immune cells in blood may help regulate neurophysiology and behavior [51]. While these correlations between blood and brain DNAm are interesting to think about, the results should not be overstated. However, if paired with in vitro or in vivo analyses where brain tissue can be studied directly, we may be able to gain stronger mechanistic insight probing questions identified in peripheral blood analyses.

A number of interesting findings appeared in both the GO and KEGG pathways analyses. These were exploratory analyses since we used the top loci by raw p-value, and not all results would withstand correction for multiple hypothesis testing (q < 0.2). Even so, pathways related to neurotransmission, immune response and inflammation were common observations across analyses for all outcomes. The KEGG term for glutamatergic synapses was enriched for both the block change and ISI change analyses (Tables 4 & 5). The dysregulation of glutamatergic neuronal functioning is implicated in the etiology of ADHD [52,53]. Its enrichment in our study for block change and ISI change outcomes suggests glutamatergic neurons may be more relevant to sustained attention and vigilance, respectively.

Among top loci associated with perseveration, pathways analysis using the KEGG database revealed enrichment for cGMP–PKG signaling as well as a number of calcium-related signaling pathways. It has been shown that cGMP–PKG signaling acts as an intrinsic modulatory system by regulating neuronal calcium levels [54]. This can be accomplished by activating cyclic nucleotide-gated channels or releasing calcium stores via ryanodine receptors [55,56]. Calcium pathways were also enriched among perseveration results when using the GO database. These findings suggest that calcium dysregulation in the brain may influence impulsivity. This claim is further supported by Altun Varmis et al., who found that increased calcium levels were associated with ADHD, as were reduced parathyroid hormone levels, which helps regulate calcium levels [57]. In a childhood ADHD EWAS conducted by Mooney et al., the top-ranked differentially methylated position, showing increased DNAm in ADHD cases compared with non-ADHD controls, was located in the promotor region of SLC7A8 [6]. This gene is involved in several functions, including response to elevated platelet cytosolic calcium levels and thyroid hormone transport.

Another KEGG pathway enriched among loci genes associated with perseveration, as well as with ISI change, was ferroptosis. This was the only pathway that was significantly enriched at the false discovery rate adjusted significance value. Ferroptosis is a highly regulated, complex, iron-dependent form of cell death that is characterized by the oxidative destruction of the cell’s lipid bilayer. Recent studies suggest that dysregulation of ferroptosis may be closely related to the pathogenesis of neurological disorders including neurodegenerative diseases, strokes and glioma [58–60]. Interestingly, iron deficiency in various brain regions, including the globus pallidus, caudate nucleus and hippocampus, was characteristic of children with ADHD according to a quantitative susceptibility mapping study conducted by Tang et al. [61]. This, partnered with our study’s findings, suggests that iron deficiency in the brain and subsequent dysregulation of the ferroptosis pathway may be associated with ADHD in children. Growing recognition by the scientific community that iron deficiency may be associated with ADHD and the findings of a recent randomized clinical controlled trial suggest that iron supplementation may reduce ADHD symptom severity in children, at least when partnered with methylphenidate treatment [62,63]. Additional research is needed to discern the true effects of iron supplementation both in isolation and in concert with methylphenidate treatment or other pharmacological interventions. Furthermore, the role of ferroptosis in this process until this point has only been implied; mechanistic studies of this process in relation to ADHD may be warranted.

Abnormal inflammatory response and subsequent immune dysregulation are implicated in the etiology of ADHD [64]. Numerous studies have linked other psychiatric illnesses to aberrant inflammatory pathways [65,66]. Additionally, anxiety and stress are important risk factors and comorbidities of ADHD and are also associated with inflammation and immune dysregulation [67–69]. Results from the present study support the idea that inflammation and immune dysregulation may play a role in the development of ADHD. The CpG site cg12603272 corresponding to the 5′UTR of ELF4, an X-linked erythroblast transformation specific transcription factor gene that plays an important role in regulating inflammation and immune response, was statistically significantly associated with the ISI change outcome. Generally, ELF4 suppresses macrophage response and regulates anti- and pro-inflammatory genes in macrophages. Deficiency in ELF4 increases inflammatory T_H17 cell responses [70]. In the context of our study, increased DNAm at this CpG site was associated with higher CPT3 scores, or a decrease in vigilance. DNAm at a given gene is often associated with transcriptional suppression of that gene [71]. We used a subset of ELEMENT participants (n = 69) who also had transcriptomic data available to quantify the associations between DNAm and expression at the genes of interest; though not statistically significant, the expected inverse relationship between DNAm and expression was observed for ZNF814 and OR6K6 genes (Table 6). For the ELF4 gene, no relationship was observed (r = 0.08; p = 0.52) so we cannot draw any inferences regarding how expression levels of this gene may be related to CPT3 performance or the etiology of ADHD. Notably, these transcriptomic data were quantified in blood and not in brain, which is the biologically relevant tissue to the outcome. Additionally, we were not able to distinguish between transcriptional variants through this analysis. The gene counts data we generated from the RNA sequencing results represent expression of any transcriptional variant of a given gene. It is entirely possible that methylation at a specific probe can affect not only gross transcriptional levels, but also the ratio of variants transcribed. This is biologically relevant, because distinct variants may be associated with different levels of transcript stability, translational efficiency or altered protein structure that could affect the gene product’s activity without significantly impacting gross expression levels themselves. At least for the example of ZNF814, according to both Gencode and RefSeq, transcriptional variants are similar in the region where the differentially methylated loci are (Supplementary Figure 2) [72,73].

Inflammatory and immune pathways were enriched for all four CPT3 outcomes. Biological process and molecular function terms related to various types of interleukins or leukocytes were present among the top GO pathways for all outcomes. Interleukins are a class of cytokines comprising glycoproteins produced by white blood cells that aid in immune response regulation. Because our EWAS used whole blood to quantify DNAm, pathways analyses were also conducted using the sensitivity models for all four outcomes, as they controlled for cell type. As noted in the final columns of Tables 4 & 5, & Supplementary Tables 5–10, several of these immune-related pathways remained enriched in the sensitivity analyses, especially for the ISI outcome. Examples include negative regulation of leukocyte degranulation, negative regulation of myeloid leukocyte-mediated immunity, and CD8⁺, α/β-T-cell lineage commitment. The MAPK signaling pathway, playing a role in inflammation and immune cell activation in the brain, is another important example of a pathway that was enriched for KEGG analyses for both block change and perseveration outcomes. One study suggests that disruption of MAPK signaling may result in autism spectrum disorder [74]. Preliminary evidence also exists supporting ADHD as a neuroinflammatory disease characterized by activation of immune cells in the brain, perturbation of the blood–brain barrier, and activation of MAPK and NF-κB signaling [75]. One EWAS assessing relationships with externalizing behavior (characterized by aggression, disobedience, impulsivity and poor emotional regulation) in children and adolescents noted significant differential methylation at ADHD-related genes and saw enrichment for pathways related to neuroinflammation. The results of their study suggest that interactions between neuroinflammation and dopaminergic pathways may underlie externalizing behaviors, such as those exhibited by children with ADHD [7]. As several of the immune/inflammation pathways identified by our study were enriched in both main and sensitivity models, we believe the results of our present research are consistent with the fount of literature supporting ADHD as a neuroinflammatory disease.

Previous research has explored the relationship between DNAm across the epigenome and ADHD (or ADHD-like symptoms). The significant loci identified in our study did not overlap with those previously reported, but these studies vary in the age of study participants, tissue used for DNAm analysis (e.g., blood, saliva), method of assessment for attention (e.g., clinical diagnosis of ADHD, neuropsychological test, self/parent-reported documentation of symptoms, etc.) and more. Studies focusing on the relationship between DNAm and neurodevelopmental outcomes, particularly ADHD symptoms and externalizing behaviors, include: 1) Mooney et al. 20202) Li et al. 2021; 3) Sun et al. 2023; 4) Neumann et al. 2020 and 5) Meijer et al. 2023 [5–8,76]. Mooney et al. conducted an EWAS of childhood ADHD using peripheral saliva DNA from 603 children (391 with ADHD and 213 without ADHD) [6]. The study found epigenome-wide significant associations between ADHD polygenic risk and DNAm at sites annotated to the promoter of genes GART and SON. The study also identified DNAm markers that were associated with clinical diagnosis of ADHD, though not via Bonferroni-corrected p-value threshold of 8.8E -08. Hypermethylation in ADHD cases relative to controls was observed at the two top-ranked dinucleotide CpG sites, cg17478313 (Δβ: 0.93%; p = 1.54E-06) and cg21609804 (Δβ: 1.38%; p = 2.82E-06). These sites are respectively located in the promotor region of SLC7A8 and the 3′-UTR of MARK2 [6].

One of the major sources of variability across ADHD EWASs is the method used for phenotyping the outcome and for quantifying methylation. The parent-, teacher- or self-rated Strengths and Difficulties Questionnaire (SDQ), for example, is a widely used screening tool for behavioral and emotional problems in children and more recently adolescents but is innately subject to bias and human error. For this reason, it is preferable to crossvalidate self-report data with parent or teacher reports, but this is often not feasible [77,78]. This was the case for the longitudinal Sun et al. study and some of the studies considered in a combined analysis conducted by Meijer et al. that used SDQ for their assessments [7,8]. Lacking in clinical ADHD diagnosis data for a sufficient sample size for use in their models, many researchers must use the hyperactivity–inattention domain of the SDQ or similar assessments to quantify ADHD and ADHD-like symptoms, and only confirm ADHD status in participants with clinical diagnosis when data is available [8,76]. While the present study in ELEMENT participants is also lacking in clinical ADHD diagnosis data, it used data from the computerized CPT3 test for its models. Using continuous outcomes such as this may provide more power to detect small changes in symptoms but cannot technically be related to ADHD without a clinical diagnosis. Li et al. included both CPT outcomes and diagnostic data to complement one another in studying the relationships with DNAm [5]. While this study is an epigenome-wide screen (followed by site-specific pyrosequencing analysis on selected candidate genes) like the other studies discussed in this section, it is distinct in that instead of using a more traditional regression model for its statistical method it employs analysis of variance and multivariate analysis of covariance models, which is likely to underestimate standard error and makes it more difficult to compare results across studies. Furthermore, its sample size is notably small (12 cases and 9 controls), thus reducing its power. To its credit, this study, along with Mooney et al., is among the first to use the EPIC array to study potential biomarkers of ADHD. This provides a greater coverage of the epigenome, as it contains >850,000 probes, whereas the more commonly cited assay for prior EWASs, Illumina Human Methylation 450K array, only contains ∼450,000 probes [5,6].

Interestingly, Sun et al. used both the EPIC and the 450K arrays, pulling data from the IMAGEN project from eight different sites across Europe [7]. The researchers used the SDQ to investigate the relationship between DNAm and internalizing behavior or externalizing behavior. In the externalizing-behavior EWAS, only one probe (cg01460382) attained epigenome-wide significance (r = 0.18; t₁₀₀₉ = 5.74; p_uncorrected = 1.26 × 10^-8; p_Bonferroni = 4.7 × 10^-3). No significant relationships between DNAm and internalizing behavior were identified. They also related DNAm to gray matter volumes, discovered DMRs associated with behavioral problems and suggested these DMRs may play a role in facilitating brain development [7]. Neumann et al. also employed data from both Illumina EPIC and 450K arrays. They conducted a prospective meta-analysis of research assessing the association between DNAm and ADHD symptoms from birth to school age (12 years old) as determined by parent-reported questionnaires. DNAm in school-age children was not associated with ADHD symptoms [76]. One potential factor that could be contributing to the lack of significant findings in these multiarray analyses is the fact that even though there is an overlap in probes across these assays, DNAm from the EPIC and 450K array are not readily comparable to one another and do not reliably replicate results on the individual probe level even within the same samples [79]. This poses a challenge to replication studies and meta-analyses such as these, but fortunately computational methods have been designed to address this issue and harmonize probes across arrays [80,81]. The research outlined by Neuman et al. included cohorts based in Europe, including Denmark, England, Finland, The Netherlands and Spain, but did not include any from Latin America, as is seen in this current analysis [76]. This also sets the present research aside from the previously discussed studies such as Mooney et al., Meijer et al. and Sun et al., whose cohort participants are predominantly of North American and European descent [6–8,13]. It is important to note that relationships between specific DNAm markers and ADHD symptoms observed in the Neumann et al. meta-analysis varied across different cohorts and time points, suggesting the role of DNAm in ADHD etiology may differ based on developmental timepoint, demographics and other risk factors of the underlying population [76].

Meijer et al. conducted a combined analysis and review of epigenome-wide DNAm and externalizing behaviors in children, adolescents and adults, with ten studies meeting their inclusion criteria [8]. Across multiple studies, this review identified several DMRs that were consistently associated with externalizing behaviors in children and adolescents and proposed ten candidate genes for ADHD: GNG7, ADAMTS2, TNXB, ATP11A, BRD2, CACNA1H, SDK2, ACOXL, FLJ30306 and ARID4A. Meijer et al. also highlighted the importance of considering the age range of participants, tissue used for DNAm quantification and methods employed to assess their externalizing behavior outcomes for appropriate interpretation of results. The studies included in their analysis measured externalizing behaviors in different ways, including the SDQ, behavioral observations and diagnostic interviews. While this combined analysis does not solely relate DNAm to ADHD diagnosis or ADHD-specific symptoms and instead uses the broader term of externalizing behavior, ADHD is one of the most commonly occurring comorbidities of externalizing behavior disorders and the findings from this study are comparable to more ADHD-specific studies [8,82]. One study in a Swedish birth cohort suggested externalizing behaviors in middle childhood were predictive of ADHD-like symptoms in early adolescence, and saw associations between ADHD-like traits and externalizing behaviors increased over time [83]. This underscores the interrelationship between these commonly concurring symptoms and disorders.

Overall, previous studies suggest that DNAm may be a promising biomarker for predicting and understanding neurodevelopmental outcomes such as ADHD symptoms and externalizing behaviors. Several proposed gene biomarkers of ADHD, as identified by Meijer et al., contain significant findings across different studies. The relevant probes annotated to these genes were extracted from the results files of this current investigation. Only probes that were identified as associated with ADHD in child or adolescent populations were extracted for comparison and compiled into document Supplemental B. While none of these results were significant on the epigenome-wide level, some raw p-values were significant (p < 0.05) and many of the probes had similar direction of effect as observed in prior studies. The probe cg15829622, for instance, was significant for both ISI change (β: -0.582; SE: 0.20; p = 0.004) and block change (β: -0.429; SE: 0.19; p = 0.024) EWASs. Not only was the direction of effect similar for both analyses, with lower DNAm associating with worse CPT3 performance (higher scores), but it was also similar to what we would expect from Goodman et al., which document and an association between lower levels of DNAm at this probe and ADHD (β: -0.96; p = 4.57E-04) [8,84].

Epigenomic research focused on ADHD to date is met with many challenges. For one, the lack of consistency in assay type utilized makes comparisons across studies difficult. As previously mentioned, the Illumina 450K and EPIC arrays, despite covering many of the same CpG sites and quantifying DNAm through the same principles, fail to produce consistent findings even when identical DNA sources are used [79–81]. This could be attributed to unequal representation of type I (containing two sequences per probe: one for methylated and the other for unmethylated versions) and type II probes (containing only one sequence, occupying half the physical space as type I probes). For this reason, one should take caution in drawing conclusions and interpreting results when making cross-study comparisons. In addition to variability between the different iterations of the Illumina Infinium platforms, there can also be batch effects across experiments of the same type and even between probes on an individual chip [16,79,81,85]. Another challenge to the field is the lack of a standard threshold for significance. This study uses a Benjamini–Hochberg method to correct for multiple comparisons with a q-value cutoff of 0.2. This is moderately stringent, with prior studies documenting significant findings based on far more lenient cutoffs and others far stricter [8,33,86]. This could be considered a weakness of our study, which is why results with a more conventional q-value cutoff of 0.05 are emphasized in our discussion.

Prior research also emphasizes the relevance of considering participant age and tissue source when interpreting results and provides insight into potential mechanisms underlying the development of these outcomes [76]. While these studies offer a starting point to uncover the associations of DNAm with ADHD etiology, more research is warranted in diverse populations with more ADHD-specific outcomes. Unlike previous studies, this research is conducted in a Hispanic population. This investigation also builds upon past research by using an ADHD-specific test completed by the participant which employs computerized quantifications of varying aspects of attention, rather than survey responses from external parties like parents or teachers. Future research may continue to explore these associations and their potential implications for diagnosis and treatment in diverse cohorts.

As with any study, ours is met with both strengths and weaknesses. Notably, this study is among the first to employ an epigenome-wide screen to identify associations between DNAm and ADHD-like symptoms in adolescence, a relevant neurodevelopmental timepoint [5–8]. To our knowledge it is also the first study to do so in a Hispanic population. Another strength is the large sample size of 517 participants. While this provides power to detect small changes in DNAm in the full sample, it is limited in its power to detect gender-specific differences. Notably, past studies have documented gender-specific associations between DNAm and ADHD-related phenotypes, but only two significant associations were identified in this study with one outcome [6,8]. The present study also utilizes the results from the CPT3 test, which quantifies symptoms related to ADHD as continuous variables but does not equate them to a diagnosis [23–25]. While using this test in our analyses impedes our ability to draw a relationship between DNAm and a clinical diagnosis for ADHD, the continuous nature of these CPT3 variables strengthens our study’s power to identify associations with small differences in test performance reflecting behavioral changes that might not signify an official diagnosis. One limitation is the use of blood leukocyte DNA for EWAS analyses instead of brain DNA, the ideal tissue for a neurological outcome. While methylation profiles in blood correlate with those of the brain at some genes, they differ at many others [34]. Notably, all CPT3 outcomes met the assumptions of linearity and were modeled using linear regression. It is possible, however, that the relationship between DNAm at some loci and outcomes is not linear and was not identified here. Nonlinear model types should be explored in future analyses to entertain the possibility of different types of relationships between DNAm at certain genes and CPT3 performance. As previously mentioned, the EPIC array has a number of technical limitations, such as the unequal representation of type I and type II probes. It also has an over-representation of cancer and other disease-related sites. It only covers a portion of the CpG sites in the entire epigenome. Computational techniques have been developed to address some of these concerns. Regardless of its challenges, the EPIC array has a larger coverage than past versions of Illumina Infinium arrays, spans the entire genome and remains to be a useful tool for epigenome-wide analyses [16,80,81,85,87]. To that extent, DNAm in any tissue type is not the only epigenetic mechanism modulating gene expression. Other factors, such as chromatin architecture, histone modification and noncoding RNA such as piRNA, also play a role and may be impacting the development of this highly complex neuropsychological disorder [4]. These mechanisms should be explored in relation to ADHD in children in future research. Regardless, the present EWAS provides a foundation for future explorations of the relationship between epigenetic mechanisms and the etiology of ADHD.

Conclusion

We conducted an EWAS of ADHD-like symptoms using blood leukocyte DNA from Mexican adolescents. We identified associations between DNAm at specific CpG sites and three out of four CPT3 variables selected to reflect ADHD symptom profiles: sustained attention (block change), vigilance (ISI change) and impulsivity (perseverations). Differential DNAm at sites located in the gene body of ZNF814 were associated with these three CPT3 outcomes. Additionally, hypermethylation at ELF4 and OR6K6 were individually associated with an increase in ISI change score (i.e., a decrease in vigilance). Across all CPT3 outcomes, we observed enrichment for pathways related to neurotransmission, immune response and inflammation among differentially methylated sites.

Epigenetics, including DNAm, has great potential to serve as biomarkers of ADHD and help unravel its underlying etiology. ADHD is a complex disorder that may be influenced by many different factors, including the social environment, diet and exposure to toxic chemicals. Future research would benefit from analysis of the joint effect of DNAm and environmental exposures on the etiology of ADHD in children. Prior to that, analyses like this current study, which analyze the effects of each predictor on ADHD etiology independent of the other, are essential. While the ELEMENT cohort was not designed to include clinical ADHD diagnostic data, analyses in similarly aged populations may consider a case–control design to best identify DNAm biomarkers of ADHD. Regardless, this research laid the groundwork for future biomarker and mechanistic studies. For example, the potential role of the ZNF814 protein, as well as specific pathways, such as those related to inflammation, immune response and neurotransmission, could be considered in in vitro or in vivo experiments focused on ADHD pathogenesis.

Summary points

• Attention-deficit/hyperactivity disorder (ADHD) is a highly prevalent neurodevelopmental disorder in children and a risk factor for numerous future health outcomes.

• The complex etiology of ADHD is not well understood, but DNA methylation (DNAm) has been suggested to play a role.

• In an epigenome-wide association study of 517 ELEMENT participants, DNAm was related to childhood performance on the Conners’ continuous performance test, third edition.

• DNAm at ZNF814 was associated with vigilance, sustained attention and impulsivity.

• ELF4 and OR6K6 DNAm was individually associated with an increase in interstimulus interval change score, or a decrease in vigilance.

• Inflammation, immune response and neurotransmission pathways were enriched among differentially methylated genes across all outcomes of Conners’ continuous performance test, third edition.

• Future in vitro and in vivo studies regarding potential ADHD etiology are needed to validate any mechanistic implications discussed in this study.

• DNAm may potentially serve as a biomarker for predicting and understanding neurodevelopmental outcomes such as externalizing behaviors and ADHD symptoms, but currently more research is warranted.

Author contributions

J Ehlinger conceived the idea for the project, performed the data analysis and wrote the manuscript. JM Goodrich supervised the project, contributed to study conception and design, and generated epigenetic data. MM Téllez-Rojo and KE Peterson designed, maintained and led the ELEMENT cohort. A Cantoral and A Mercado-García acquired cohort data. DJ Watkins contributed expertise to interpret data. MM Téllez-Rojo, KE Peterson, DC Dolinoy and JM Goodrich acquired funding for the study. All authors contributed to the drafting of the manuscript and approve the final version.

Writing disclosure

No writing assistance was utilized in the production of this manuscript.

Ethical conduct of research

Researchers affirm that all research protocols were approved by the institutional review boards of the University of Michigan and the Mexico National Institute of Public Health.

They also attest that both informed consent and participant assent were obtained from the parents of all participants.

Data sharing statement

Data (epigenetic and demographic) from the ELEMENT study are available through the NIH Human Health Exposure Analysis Resource data repository (doi: 10.36043/1431_392, 10.36043/1431_327 and 10.36043/1431_393). Additional data are available upon request from the study team.

Acknowledgments

We would like to acknowledge the contributions of and express our deepest gratitude to the American British Cowdray Medical Center in Mexico for graciously allowing us to use their research facilities making this work possible. We would also like to gratefully acknowledge the tireless work of the ELEMENT research staff for their persistent work in data collection, sample preparation and experimental proceedings. We recognize the Bioinformatics Core and the Advanced Genomics Core of the University of Michigan Medical School’s Biomedical Research Core Facilities in their support of this research. Lastly, we extend our appreciation to all the cohort participants, for without their contribution none of this research would have been possible.

Supplementary data

To view the supplementary data that accompany this paper please visit the journal website at: www.tandfonline.com/doi/suppl/10.2217/epi-2023-0169

Financial disclosure

This ELEMENT study received funding from the US Environmental Protection Agency (grant no. RD83543601) and National Institute of Environmental Health Sciences (award nos. P01 ES02284401, R01 ES007821, R01 ES014930, R01 ES013744, P30 ES017885, 1U2C ES026553, K01 HL151673 and R35 ES031686). This study was also supported and partially funded by the National Institute of Public Health/Ministry of Health of Mexico. J Ehlinger is also supported by National Institute of Environmental Health Sciences under the Environmental Toxicology and Epidemiology training grant (award no. T32 ES007062). The authors have no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.

Competing interests disclosure

The authors have no competing interests or relevant affiliations with any organization or entity with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.