Peripheral DNA Methylation of Cortisol- and Serotonin-Related Genes Predicts Hippocampal Volume in a Pediatric Population

Abstract

Background

Hippocampal volume increases throughout early development and is an important indicator of cognitive abilities and mental health. However, hippocampal development is highly vulnerable to exposures during development, as seen by smaller hippocampal volume and differential epigenetic programming in genes implicated in mental health. However, few studies have investigated hippocampal volume in relation to the peripheral epigenome across development, and even less is known about potential genetic moderators. Therefore, in this study, we explored relationships between hippocampal volume and peripheral DNA methylation of mental health–related genes, specifically NR3C1, FKBP5, and SLC6A4, throughout early development and whether these associations were moderated by age or genotype.

Methods

Bilateral hippocampal volume was computed from T2-weighted images through FreeSurfer, and DNA methylation was measured from saliva using the Illumina MethylationEPIC microarray in a pediatric population (N = 248, females = 112, mean_age = 5.13 years, SD_age = 3.60 years).

Results

Multiple linear regression and bootstrapping analyses revealed that DNA methylation of NR3C1, FKBP5, and SLC6A4 was associated with hippocampal volume and that these relationships were moderated by age and gene-specific variants.

Conclusions

These findings support the validity of peripheral DNA methylation profiles for indirectly assessing hippocampal volume and development and underscore the importance of genotype and age considerations in research. Therefore, peripheral epigenetic profiles may be a promising avenue for investigating the impacts of early-life stress on brain structure and subsequent mental health outcomes.

Plain Language Summary

This study explores associations between DNA methylation of mental health–related genes from saliva samples and hippocampal brain volume in children. The findings suggest that DNA methylation across 3 key genes is linked to hippocampal size, with age and genetic variations playing a role in how strong these associations are. Cortisol-related genes, NR3C1 and FKBP5, are stronger predictors in early childhood, while the serotonin transporter, SLC6A4, is stronger in later childhood. These results highlight the importance of studying peripheral epigenetic profiles across development and their relation to neurodevelopment, which may offer insights for personalized mental health treatments.

Plain Language Summary

This study explores associations between DNA methylation of mental health–related genes from saliva samples and hippocampal brain volume in children. The findings suggest that DNA methylation across 3 key genes is linked to hippocampal size, with age and genetic variations playing a role in how strong these associations are. Cortisol-related genes, NR3C1 and FKBP5, are stronger predictors in early childhood, while the serotonin transporter, SLC6A4, is stronger in later childhood. These results highlight the importance of studying peripheral epigenetic profiles across development and their relation to neurodevelopment, which may offer insights for personalized mental health treatments.

The hippocampus is crucial for learning, memory, and the hypothalamic-pituitary-adrenal (HPA) axis stress response. It undergoes significant changes during development, including neurogenesis, dendritic and axonal growth, and synaptic formation and pruning (1). Reduced hippocampal volume is consistently found across adult and pediatric clinical populations, such as those with major depressive disorder (2,3) and posttraumatic stress disorder (4,5). Notably, the hippocampus is highly susceptible to early-life adversity due to its high density of glucocorticoid receptors (NR3C1), FK506-binding protein 51 cochaperones (FKBP5), and serotonin transporters (SLC6A4), all of which are vital for stress regulation (6, 7, 8). Epigenetic processes likely mediate the impact of early experiences on hippocampal development, with subsequent differential gene expression leading to HPA axis dysfunction and reduced hippocampal volume, both of which have been linked to psychiatric disorders (2, 3, 4, 5,9).

Epigenetics involves gene regulation mechanisms independent of DNA sequence, with DNA methylation (DNAm) being the most studied in relation to exposures and mental health in adults (10) and children (11,12). Pioneering rodent studies demonstrated that maternal care variations can modify DNAm and NR3C1 expression in hippocampal tissue (13), a finding extended to human studies using postmortem brain tissue and peripheral samples by our group and others (14,15). Additional animal and human studies support the role of epigenetic action on FKBP5 and SLC6A4 in response to early experiences (15, 16, 17). Notably, human epigenetic studies have often used peripheral samples, raising scientific interest in how well the peripheral epigenome reflects central nervous system processes.

In neuroimaging epigenetics, peripheral samples have shown mixed results regarding correlations between hippocampal volume and DNAm of NR3C1, FKBP5, and SLC6A4. A study found positive associations between blood DNAm of NR3C1 promoter sites and hippocampal subfield volumes in adults and patients with major depressive disorder (18), but this was not replicated in buccal or saliva samples (19). Similarly, FKBP5 intron 7 CpG sites did not correlate significantly with hippocampal volume in blood, buccal, or saliva samples (19,20). Conflicting results were also found for SLC6A4 promoter region CpG sites in relation to hippocampal volume (21,22). A recent epigenome-wide association study (EWAS) in adults linked hippocampal volume to 2 sites (23), neither of which are the current study’s candidate genes. No studies have yet assessed peripheral DNAm and hippocampal volume in pediatric cohorts; only 2 have done so in adolescents, and both showed modest associations from blood (24) and saliva (25) samples. These findings highlight the need to consider development when linking DNAm of NR3C1, FKBP5, and SLC6A4 with hippocampal volume. Past studies were limited by focusing on few CpG sites and lacking genetic moderation considerations. While saliva samples correlate most strongly with brain DNAm (26, 27, 28), variability in peripheral sample types may contribute to inconsistent results.

Genetic context can significantly influence epigenetic patterns and phenotypic expressions. Genetic variations within genes related to neuroplasticity, HPA axis, and dopaminergic and serotonergic systems can moderate the associations between psychiatric symptoms (29, 30, 31, 32) and early-life exposures with DNAm (33,34). Recent neuroimaging multi-omic studies emphasize the need to model both genotype and DNAm when predicting brain development (35), activity (36,37), structure (24,38), and white matter lesions (39). These findings highlight the necessity of considering genetic variation when examining epigenetic associations with brain endophenotypes.

During early development, both the hippocampus and epigenome undergo significant changes and are highly susceptible to environmental influences. NR3C1, FKBP5, and SLC6A4 are among the most studied genes in relation to early-life adversity and psychiatric symptoms, thereby meriting specific consideration. This study addresses the current literature gap by assessing the relationships between DNAm of NR3C1, FKBP5, and SLC6A4 and hippocampal volume across early development. Using data from the ECHO (Environmental influences on Child Health Outcomes) program, we investigated 1) associations between DNAm of these genes at the gene and CpG levels with hippocampal volume across early development, 2) age as a moderator of these relationships, and 3) genetic variants within these genes as moderators of the epigenome-brain relationships.

Methods and Materials

Parent Study

This study used data from the ECHO program, a consortium of 69 pediatric cohort studies that has been collecting data under a common protocol since 2019 (40). Single and cohort-specific institutional review boards oversaw human subject activities, with participants providing written informed consent. Eligibility criteria for the parent study included mothers >18 years old, term gestation (37–41 weeks), healthy singleton pregnancy, no uncontrolled medical conditions, no history of major psychiatric illness, English speaking, and consent to infant brain imaging and a study that is longitudinal in nature. Infants had no significant congenital anomalies, neurological trauma, or disorder. Inclusion criteria for this subset required participants to provide a saliva sample and a magnetic resonance imaging (MRI) scan within 365 days of each other. Written consent was collected from the parent or legal guardian.

Demographics

Demographic information was parent reported for all participants (N = 248) (Table 1). Participants were excluded if the days between saliva sample and MRI scan exceeded 365 (mean = 28.56 days, SD = 70.57). Socioeconomic status was calculated using the Hollingshead Four Factor Index of Social Status (41), with maternal and paternal education scores multiplied by 3 and then averaged, resulting in weighted average socioeconomic status scores.

Table 1.

Sample Descriptive Statistics

	Values
Age
Mean ± SD, Years	5.13 ± 3.6
Range	2 months–14 years
Sex
Female	45%
Male	55%
Race
American Indian or Alaska Native	0.4%
Asian	2.0%
Asian Indian	0.8%
Black or African American	7.3%
More than one	13.7%
Native Hawaiian or Pacific Islander	0.4%
Other, unknown, declined, or missing	6.9%
White	68.5%
Ethnicity
Non-Hispanic/Latino	74.6%
Hispanic/Latino	21.0%
Declined or missing	4.4%

Saliva Collection and DNA Isolation

Saliva was collected using Oragene kits (DNA Genotek) during laboratory visits. DNA was extracted with the supplier’s isolation kit and assessed for purity and yield using a NanoDrop ND-1000 (Thermo Fisher Scientific).

Genotype Array and Data Preprocessing

Genotyping was conducted at the Translational Genomics Research Institute using a Multi-Ethnic Global Array (>1.7 million markers) on an Illumina iScanSystem. Data preprocessing and quality control were done using PLINK version 1.9 (42), applying thresholds for single nucleotide polymorphism (SNP) genotyping rate (≥95%), minor allele frequency (≥5%), Hardy-Weinberg equilibrium (p ≥ 1.0 × 10⁻⁵), and sample genotyping rate (≥95%). Sex mismatch analysis and relatedness detection were performed, and principal component analyses (PCAs) were conducted to remove outliers. VCF files were used for imputation via the TOPMed Imputation Server. The first 10 ancestry principal components (PCs) and eigenvalues were calculated using PLINK (--pca) with default minor allele frequency (--maf) and call rate (--geno) filters. Based on the eigenvalue scree plot, the first 3 PCs were used in future analyses and subsequently confirmed as providing the best model fit compared with 4 or more components (Akaike information criterion = 3207.4, adjusted R² = 0.6394).

DNAm Array and Data Preprocessing

Bisulfite conversion was completed using the EZ-96 DNAm kit (Zymo Research). DNAm levels were measured using an Illumina Infinium MethylationEPIC BeadChip (850K) array. Data preprocessing involved using the minfi package in R, excluding cross-reactive and sex chromosome probes, and performing quality control measures including quantile normalization (43). The proportion of epithelial cells per sample was estimated using the robust partial correlation method from EpiDISH (44).

MRI Scans and Structural Measurement

MRI was conducted by collaborators at Brown University, and detailed methods can be found in Deoni et al. (45). Processing of MRI images was performed using FreeSurfer software, utilizing infant FreeSurfer atlases (46). FreeSurfer then delineated 33 distinct brain regions per hemisphere. Regional brain volumes were reported in voxels (mm³).

Statistical Analyses

PC Analysis

M-values of CpG sites annotated to each candidate gene were used to calculate PCs from PCA, following recommendations (47). M-values, the log₂ ratio of beta values, were used to resolve heteroscedasticity in highly methylated or unmethylated probes, stabilizing variance. M-values also enhance detection rates and true positive rates. PCA, performed using SPSS (IBM Corp.), used the first PC (PC1) in analyses. PCs capture data variance more effectively than summed or weighted DNAm scores (48), especially with correlated CpG sites, by transforming them into uncorrelated PCs (49). Using PCs reduces the number of statistical tests needed, thereby minimizing type 1 error. PCA is well established in DNAm studies and commonly used in pediatric cohorts (14,50,51).

Multiple Linear Regressions

The outcome variable was computed by averaging right and left hippocampal volume. The PC1 of each candidate gene was the predictor variable, including age at brain scan, sex, days between genetic collection and brain scan, cell count, batch, and array position as covariates (model 1, N = 248). We also tested subsequent models with the addition of 3 ancestry components (model 2, n = 219) and socioeconomic status (model 3, n = 180). Lastly, we assessed individual CpG sites across all 3 candidate genes as predictors of hippocampal volume in model 3 (false discovery rate p value correction; q < .05). While “predict” is used here as standard regression language, these analyses do not imply causality.

Genotype Interactions

Due to long-recognized difficulties in detecting interactions among continuous variables in moderated multiple regression analysis (52), interactions were tested with model 1 to retain the largest sample size. SNPs assessed can be found in Table S1 (24,53, 54, 55, 56, 57). Multiple linear regressions analyzed interactions between CpG sites and SNPs for each gene, with covariates as before. Due to the smaller sample size available with DNAm, MRI, and genotype data, we used a more lenient false discovery rate threshold common in the literature (q < .1) (58, 59, 60).

Bootstrap Resampling

To validate our sampling distribution, a bootstrap analysis was performed on the correlations between hippocampal volume and the PC1 of each gene using the boot package in R. The simulation was repeated 10,000 times to obtain stable sampling distribution estimates, and confidence intervals were calculated via the percentile method.

Correlations Between Saliva and Brain DNAm

Correlations were performed between DNAm beta values from this study and those of publicly available brain tissue data for matching CpG sites. Brain DNAm values were obtained from the Allen Brain Atlas BrainSpan dataset, a publicly available dataset based on the Infinium HumanMethylation450 BeadChip. The BrainSpan dataset provided DNAm values from 16 brain regions (11 cortex regions, hippocampus, amygdala, thalamus, striatum, and cerebellum) across participants (n = 177 samples from 16 individuals, mean_age = 8.35 years, SD_age = 10.0 years). To conduct correlation analyses, we calculated the average DNAm value across participants per CpG site annotated to genes of interest. Separate correlation analyses were performed for whole brain (n = 16, mean_age = 8.35 years, SD_age = 10.0 years), hippocampus only (n = 10, mean_age = 5.86 years, SD_age = 7.03), and nonhippocampal brain regions (n = 16, mean_age = 8.35 years, SD_age = 10.0 years).

Results

PCA of DNAm Across NR3C1, FKBP5, and SLC6A4

The descriptions and locations of CpG sites retained in PCA are detailed in Table S2. CpG sites within and outside the promoter regions (e.g., islands, shores, shelves, open seas) strongly contributed to the PC1 of each gene (Table S3), demonstrating high correlation across various CpG types.

NR3C1, FKBP5, and SLC6A4 DNAm Predicting Hippocampal Volume

PC1 DNAm Predicting Hippocampal Volume

Across all models, all 3 candidate gene PC1s significantly predicted bilateral hippocampal volume (Table 2 and Figure 1). The full model (model 3) produced the highest effect sizes and best model fit across all genes, as evidenced by the highest beta and R² values (Table 2). Bootstrapping results demonstrated stability and accuracy, with all correlation values falling within the 95% CI estimated distribution (Table 3), supporting the reliability of these findings. To further validate our findings, we used lateral ventricle volume as a control analysis using the full model, revealing that the PC1s of NR3C1 (b = −7.38, p = .44), FKBP5 (b = 5.70, p = .51), and SLC6A4 (b = 7.29, p = .43) were not significant predictors, underscoring the specificity of these associations with hippocampal volume.

Table 2.

Model Fits

Model	n	Parameters	Beta	p	AIC	Adjusted R2
NR3C1
1	248	7	−6.56	2.84 × 10−5	2624.26	0.63
2	219	10	−7.18	2.60 × 10−4	2623.19	0.66
3	180	11	−6.87	1.10 × 10−4	2612.66	0.67
FKBP5
1	248	7	−3.26	.025	2615.83	0.61
2	219	10	−4.27	.008	2612.96	0.64
3	180	11	−3.72	.022	2623.48	0.65
SLC6A4
1	248	7	5.80	6.76 × 10−5	2618.98	0.63
2	219	10	5.81	1.65 × 10−4	2615.83	0.65
3	180	11	5.98	4.44 × 10−4	2615.58	0.66

Model 1 covariates include age, sex, days between genetic collection and brain scan, cell count, batch location, and batch position. Model 2 covariates include those of model 1 and the first 3 ancestry principal components. Model 3 covariates include those of model 2 and socioeconomic status. Akaike information criterion (AIC) is reported for n = 180.

Figure 1.

DNA methylation across NR3C1, FKBP5, and SLC6A4 predicts bilateral hippocampal volume in a pediatric cohort. The PC1 of (A, D)NR3C1 (b = −6.87, p < .001), (B, E)FKBP5 (b = −3.72, p = .022), and (C, F)SLC6A4 (b = 5.98, p < .001) significantly predict bilateral hippocampal volume. (A–C) depict data points, beta coefficients, and p values as corrected by full model covariates; (D–F) depict raw data points, beta coefficients, and p values with no correction from covariates. PC1, first principal component.

Table 3.

Results of Bootstrap Resampling

Gene	Original	95% CI
NR3C1	−0.72	−0.78 to −0.66
FKBP5	−0.68	−0.74 to −0.61
SLC6A4	0.72	0.65 to 0.77

CpG Site DNAm Predicting Hippocampal Volume

Based on PC1 best model fit, we tested individual CpG site DNAm as predictors using model 3. Twenty of 113 (17.7%) sites across NR3C1, 10 of 58 (17.2%) sites across FKBP5, and 8 of 35 (22.9%) sites across SLC6A4 significantly predicted bilateral hippocampal volume (qs < .05) (Table 4). Of these 38 CpG sites, 26 (68.4%) were open sea, 7 (18.4%) were shores, and 5 (13.2%) were islands.

Table 4.

CpG Site DNA Methylation Predicting Hippocampal Volume

Gene	CpG Site	CpG Location	Genomic Location	Beta Direction	abvB	p	q
NR3C1	PC1			−	6.87	1.10 × 10−4
	cg21702128	Island	chr5 142784721	+	5.99	.007	.040
	cg18718518	South Shore	chr5 142785455	−	5.84	3.77 × 10−4	.004
	cg07197341	Open Sea	chr5 142916018	+	4.64	3.80 × 10−8	2.15 × 10−6
	cg12969488	North Shore	chr5 142780984	−	4.63	4.68 × 10−6	8.81 × 10−5
	cg08320082	Open Sea	chr5 142889494	+	4.56	2.18 × 10−6	5.20 × 10−5
	cg17779063	Open Sea	chr5 142922486	+	3.97	1.63 × 10−6	5.20 × 10−5
	cg24052866	Open Sea	chr5 142727470	+	3.58	.002	.015
	cg13514002	Open Sea	chr5 142677292	+	3.33	3.43 × 10−4	.004
	cg12888360	Open Sea	chr5 142732502	−	3.30	1.88 × 10−8	2.12 × 10−6
	cg17860381	Island	chr5 142783569	+	3.03	.002	.015
	cg23776787	Open Sea	chr5 142814315	+	3.02	1.50 × 10−4	0.002
	cg15115787	Open Sea	chr5 142730701	−	2.79	2.30 × 10−6	5.20 × 10−5
	cg20728768	Open Sea	chr5 142696594	−	2.74	3.00 × 10−5	4.84 × 10−4
	cg15910486	Island	chr5 142783621	+	2.50	.005	.032
	cg23430507	Open Sea	chr5 142798375	+	2.44	.002	.015
	cg01751279	Open Sea	chr5 142793924	+	2.35	.002	.015
	cg05900547	Open Sea	chr5 142769791	−	2.26	2.55 × 10−4	.003
	cg07733851	North Shore	chr5 142781498	−	2.25	.006	.038
	cg23484741	Open Sea	chr5 142936868	+	2.09	.005	.032
	cg25708981	Open Sea	chr5 142697868	−	1.78	.001	.005
FKBP5	cg22812853	Open Sea	chr6 35512865	+	6.50	1.15 × 10−13	6.67 × 10−12
	cg03245912	South Shore	chr6 35657040	+	5.56	3.56 × 10−4	.003
	cg01731192	Open Sea	chr6 35511360	+	5.48	7.68 × 10−8	2.23 × 10−6
	PC1			−	3.72	.022
	cg15929276	Open Sea	chr6 35687457	+	3.64	5.79 × 10−6	6.72 × 10−5
	cg22363520	Open Sea	chr6 35558488	+	3.62	1.01 × 10−4	.001
	cg13344434	Open Sea	chr6 35570573	−	2.95	1.03 × 10−5	9.96 × 10−5
	cg09318204	Open Sea	chr6 35511434	+	2.88	4.39 × 10−7	6.37 × 10−6
	cg14339974	Open Sea	chr6 35687310	+	2.79	3.86 × 10−7	6.37 × 10−6
	cg16912838	Open Sea	chr6 35551624	−	1.97	.001	.004
	cg07696519	Open Sea	chr6 35619165	−	1.37	.004	.020
SLC6A4	PC1			+	5.98	4.44 × 10−4
	cg26438554	Island	chr17 28562733	−	5.59	.002	.012
	cg20209182	Open Sea	chr17 28530849	+	5.34	2.08 × 10−6	2.43 × 10−5
	cg06373684	Island	chr17 28562751	−	4.84	.004	.015
	cg01991100	Open Sea	chr17 28555935	+	3.94	3.35 × 10−4	0.002
	cg14352032	South Shore	chr17 28564834	+	3.64	2.07 × 10−6	2.43 × 10−5
	cg09921370	Open Sea	chr17 28555315	+	3.17	2.48 × 10−4	.002
	cg05951817	North Shore	chr17 28562142	+	3.11	4.49 × 10−8	1.57 × 10−6
	cg22584138	North Shore	chr17 28562220	+	2.30	.002	.012

CpG sites that did not pass false discovery rate correction at q < .05 are not reported. Covariates include age, sex, days between genetic collection and brain scan, cell count, batch, position, first 3 ancestry PCs, and socioeconomic status.

abvB, absolute value of beta; chr, chromosome; PC, principal component.

Sex, Age, and Genotype as Moderators of DNAm and Hippocampal Volume

Sex and Age Interactions

Due to sex and age being additional significant predictors of hippocampal volume, we explored models with sex and age interaction terms. To optimize interaction effect detection in multiple regression analyses, model 1 was utilized to retain the largest sample size (52). Sex did not significantly interact with NR3C1, FKBP5, or SLC6A4 PC1 to predict hippocampal volume (Table 5). However, age did significantly interact with NR3C1, FKBP5, and SLC6A4 PC1 in predicting hippocampal volume (Table 5). Next, we compared effect sizes of DNAm PC1 predicting hippocampal volume within 2 age groups based on previous literature (61), highlighting differing epigenetic malleability within these ages (younger: 0–3 years old [n = 81] and older: 3–14 years old [n = 167]). Younger children had larger effect sizes for NR3C1 (b = −1.96) and FKBP5 (b = −3.58) PC1s than older children (NR3C1 b = −0.58, FKBP5 b = −1.05) (Figure 2). In contrast, older children had a larger effect size for SLC6A4 PC1 (b = 1.94) than younger children (b = 1.46) (Figure 2).

Table 5.

Hippocampal Volume and DNA Methylation Interactions With Age and Sex

	NR3C1		FKBP5		SLC6A4
	Beta	p	Beta	p	Beta	p
Age Interaction
PC1	−5.813	2.14 × 10−5a	−4.755	2.14 × 10−4a	4.786	2.35 × 10−4a
Age (in years)	0.823	<2 × 10−16a	0.852	<2 × 10−16a	0.832	<2 × 10−16a
PC1 × age	0.602	5.72 × 10−16a	0.595	3.09 × 10−16a	−0.571	1.43 × 10−13a
Sex Interaction
PC1	−6.588	2.98 × 10−5a	−3.355	.026a	5.811	6.84 × 10−5a
Sex	1.809	5.02 × 10−5a	1.859	4.87 × 10−5a	1.714	1.30 × 10−4a
PC1 × sex	0.085	.853	0.113	.806	0.210	.659

DNA methylation is represented by first principal components (PC1). Covariates include age at scan, sex, days between genetic collection and brain scan, cell count, batch, and position. Interactions reported with age as a continuous variable.

^aIndicates significant p value (< .05).

Figure 2.

Age moderates the relationship between DNAm and bilateral hippocampal volume. DNAm across (A)NR3C1 and (B)FKBP5 is a stronger predictor of bilateral hippocampal volume at ages 0 to 3 years old (NR3C1 b = −1.96, p = .41; FKBP5 b = −3.58, p = .20) than at ages 3 to 14 years (NR3C1 b = −0.58, p = .72; FKBP5 b = −1.05, p = .41), whereas (C) DNAm across SLC6A4 is a stronger predictor of bilateral hippocampal volume in ages 3 to 14 years old (b = 1.94, p = .19) than ages 0 to 3 years old (b = 1.46, p = .50). Beta and p values included reflect results of model interaction term. DNAm, DNA methylation; PC1, first principal component.

SNP Interactions

For major allele homozygotes versus minor allele carrier comparisons, those with the homozygous major allele (TT) of SLC6A4 SNP rs6354 had significantly larger hippocampal volumes than minor allele carriers (GT/GG). There were no other significant main effects of SNP group on hippocampal volume. We found a significant interaction between NR3C1 PC1 and rs6189/90 (b = −1.86, p = .027) (Figure 3A), such that carriers of the minor allele (TC/TT) had a stronger negative association between DNAm and hippocampal volume than homozygous major allele carriers (CC). There were no significant interactions between FKBP5 or SLC6A4 PC1s and SNPs in predicting hippocampal volume. At an uncorrected p value threshold, 19 NR3C1 CpG sites, 32 FKBP5 sites, and 3 SLC6A4 sites were significantly moderated by genotype, 2 of which passed false discovery rate 0.1 correction (Table 6; an example is visualized in Figure 3B).

Figure 3.

Genotype moderates the relationship between DNAm and hippocampal volume. (A) The association between the NR3C1 DNAm PC1 and bilateral hippocampal volume depends on rs6189/90 (b = −1.86, p = .027), such that carriers of the minor allele (TC/TT) have a stronger negative association than homozygous carriers of major allele (CC). (B) The association between FKBP5 cg21789597 DNAm and hippocampal volume depends on rs9394309 (b = −3.11, q = 0.04), such that carriers of the minor allele (GA/AA) have a negative association, and homozygous carriers of the major allele (AA) have a positive association. Beta and p values included reflect results of the model interaction term. DNAm, DNA methylation; PC1, first principal component.

Table 6.

SNP × CpG Site Interactions Predicting Hippocampal Volume

Gene	SNP	CpG	Location	CpG Location	Distance	Direction	Strand	q
FKBP5	rs9394309	cg21789597	Open Sea	chr6 35633557	11,676	Upstream	cis	.03588
	rs4713916	cg21789597	Open Sea	chr6 35633557	36,325	Downstream	cis	.04786

Covariates include age at scan, sex, days between genetic collection and brain scan, cell count, batch, and position. Table includes only interactions that passed false discovery rate multiple corrections at q < .1.

SNP, single nucleotide polymorphism.

Correlations Between Buccal DNAm and Brain DNAm

DNAm measured from saliva samples in this study were highly correlated with DNAm measured in brain tissue samples in the Allen Brain Atlas BrainSpan dataset across 16 brain regions for NR3C1 (r = 0.91, p < .001), FKBP5 (r = 0.94, p < .001), SLC6A4 (r = 0.89, p < .001), and for all genes combined (r = 0.91, p < .001) (Table 7 and Figure 4). We also correlated DNAm from saliva with only hippocampal tissue or only nonhippocampal tissue, which did not change the results (Table 7 and Figure 4).

Table 7.

DNA Methylation Correlation in Brain vs. Saliva

Gene	No. of CpGs	Whole Brain		Hippocampus Only		Non-Hippocampus
		r	p	r	p	r	p
NR3C1	46	0.91	<2.2 × 10−16a	0.91	<2.2 × 10−16a	0.91	<2.2 × 10−16a
FKBP5	33	0.94	<2.2 × 10−16a	0.95	<2.2 × 10−16a	0.94	2.37 × 10−16a
SLC6A4	19	0.89	4.17 × 10−7a	0.88	6.48 × 10−7a	0.89	4.07 × 10−7a
Overall	98	0.91	<2.2 × 10−16a	0.92	<2.2 × 10−16a	0.91	<2.2 × 10−16a

^aIndicates significant p value (< .05).

Figure 4.

DNAm in saliva samples is correlated with DNAm in brain tissue. DNAm of saliva samples in the current study was highly correlated with DNAm in brain tissue samples obtained from the Allen Brain Atlas BrainSpan dataset across 16 brain regions. High correlations were present for all 3 candidate genes with DNAm measured in (A) whole brain (r = 0.91, p < .001), (B) hippocampus only (r = 0.92, p < .001), and (C) nonhippocampal regions only (r = 0.91, p < .001). Data points represent average DNAm values for individual CpG sites. DNAm, DNA methylation.

Discussion

This study contributes to neuroimaging epigenetics by showing that DNAm across NR3C1, FKBP5, and SLC6A4 from saliva is linked to hippocampal volume in a healthy pediatric population. Age moderated these relationships, with NR3C1 and FKBP5 DNAm being stronger predictors in early childhood, while SLC6A4 DNAm was stronger in later childhood. Genetic variants also moderated these relationships. Most research on peripheral epigenome and brain structure/function relationships has been conducted with adults (62), but exploring early development is key to understanding how peripheral epigenetic profiles are related to brain and behavior in pediatric populations (63).

Our PCA-based gene-wide DNAm score approach diverges from the literature. We found that NR3C1 and SLC6A4 DNAm PC1s were stronger predictors of hippocampal volume than individual CpG sites. For FKBP5, an open sea CpG site was the strongest predictor. This highlights the importance of using gene-wide DNAm and examining nonpromoter regions. Previous studies in adults on NR3C1 or SLC6A4 focused mainly on CpG sites near promoters (18,19,21,22), or in FKBP5 intron 7 (19,20), with modest results. Similarly, adolescent studies are limited in the number of CpG sites measured across SLC6A4 and FKBP5 and have yielded only modest results (25). About half of these studies also examined the average across a few sites (19,21,24). Alternatively, our approach of using PCA to assess many CpG sites reduces type 1 error and captures gene DNAm variance while also minimizing information loss. This method is robust for predicting brain structure, and we previously reported strong associations between gene PC1 and behavior, physiology, and early-life experiences (14,50,51). However, PCA limits biological interpretation of results.

To date, only one study has employed an EWAS to investigate the relationship between hippocampal volume and DNAm. This study used blood samples from adults and found no significant associations with CpG sites in the NR3C1, FKBP5, or SLC6A4 genes (23). While the EWAS approach is valuable for identifying novel, hypothesis-independent, differentially methylated regions, candidate gene approaches provide targeted analysis of specific genes of interest, which may help detect associations that could be overlooked due to the stringent p value corrections required in EWAS. In this context, a comprehensive epigenome approach may lack specificity because it evaluates associations that involve CpG sites located near genes that may not be expressed in the hippocampus. Moreover, similar to the correlation that has been observed in genotype SNP and RNA expression data (64,65), DNAm at adjacent CpG sites tends to be highly correlated (66), which presents challenges in applying unbiased p value correction methods effectively.

Historically, CpG island sites within promoter regions have been the primary focus of exposure- and behavioral-epigenetic studies, likely due to their well-characterized effects on gene transcription (67). For example, a recent meta-analysis on the effects of early-life stress and NR3C1 DNAm in humans reported that the number of assayed CpGs ranged from <10 (5 records; 15%) to more than 50 (2 records; 6%), with the majority of research reporting on a pool of 10 to 50 CpG sites of the NR3C1 gene (9 studies; 26%) (68). However, CpG sites outside of islands, such as shore, shelf, and open sea sites, are gaining attention for their regulatory roles (69,70). For example, shore CpG sites may play gene regulation roles similar to those played by island sites (70). In the context of development, CpG sites with increased and decreased age-differential DNAm are found not only in islands but also in open seas, shelfs, and shores (69). A more recent study found that age-differential DNAm was not random; CpGs with decreasing DNAm were enriched in gene bodies and enhancers and were annotated to genes enriched in immune-developmental functions. In contrast, CpGs with increasing DNAm were enriched in promoter regions and annotated to genes enriched in neurodevelopmental functions (71). Therefore, it is likely that DNAm at CpGs outside of islands is important for regulating gene transcription and should be investigated with the exposure or phenotype of interest. Results from the current study demonstrate that the majority of predictive CpG sites were open sea sites. Such findings highlight the need for further characterization of the functional consequences of DNAm at all CpG site location types.

Evidence from both rodent (72) and human (73) literature points to a hyporesponsive period in the HPA axis during early life, resulting in lower cortisol. In contrast, peak serotonin levels occur during the first 2 years of life in humans (74). Furthermore, the HPA axis is hypersensitive to exposures during the first 1000 days of life (75), whereas there is increasing evidence of heightened serotonergic sensitivity during adolescence (76,77). Interestingly, results from this study mirror these developmental stages, such that saliva-derived DNAm of NR3C1 and FKBP5 was a stronger predictor of hippocampal volume in early life, during the hypocortisol/hypersensitive stage, while SLC6A4 DNAm was a stronger predictor in later development, during the hyposerotonin/potentially hypersensitive period. While the reason for these developmental interactions cannot be ascertained by this study, it is interesting to note that correlations between the peripheral epigenome and brain metrics may be spatially and temporally dependent, with stronger relationships occurring when the relative system is hypoactive and more sensitive to exposures. Recognizing age-dependent relationships between the peripheral epigenome and brain traits is crucial for prospective studies of epigenetics and mental health in children and adolescents, which still lag behind adult studies (78). Throughout development, differing brain networks that subserve social, emotional, and cognitive development undergo increased plasticity that is sensitive to social exposures (79). Thus, mapping which gene families may best serve as peripheral biomarkers for brain development at specific periods will be an important pursuit. Replication of these results in an independent cohort and further exploration of the epigenome and brain regions involved in mental health could lead to improved understanding of which systems may be more vulnerable across specific developmental windows.

Interindividual DNAm variation is influenced by genetic variation (80, 81, 82, 83, 84), and DNAm plays a role in genetic regulatory mechanisms (81,85, 86, 87, 88, 89, 90). Thus, incorporating genetic variants in neuroimaging epigenetic studies is crucial for understanding the peripheral epigenome. For example, variations across NR3C1, FKBP5, and SLC6A4 appear to moderate the effects of early-life stress on gene expression, DNAm, brain structure and function, and behavior, such that variation in genotype can convey increased vulnerability or resiliency (91, 92, 93, 94, 95). The results of this study are consistent with this view, suggesting that HPA gene variants moderate the relationship between DNAm and hippocampal volume. One open sea site, cg21789597, significantly interacted with 2 FKBP5 SNPs, again highlighting the important role of nonisland CpG sites. NR3C1 SNP rs6189/90 had the only significant interaction with PC1 but had the least diverse genotype sample. Results also replicated similar adult findings, such that SLC6A4 SNPs did not moderate DNAm predictors of hippocampal volume (22). Taken together, this body of evidence highlights the complexity and importance of underlying genetic variance when elucidating pathways between exposures, DNAm, brain metrics, and mental health.

Structural MRI alone does not reveal transcripts, protein levels, or cellular morphology, limiting insight into the pathways that link DNAm of candidate genes with hippocampal volume. MRI-detected volumetric changes can stem from glial or endothelial cell proliferation, neurogenesis, alterations in cellular size, dendritic spine size/density, axonal remodeling, or changes in extracellular space (96,97). Few animal studies have assessed these mechanisms and have focused instead on spine density and markers for neurons and astrocytes in the mouse hippocampus (97, 98, 99). However, gray matter volume changes are more complex, involving nuclear volume, cell number, and spatial clustering characteristics (100). A recent study across 2 large adult cohorts showed that genetic variation (NR3C1 rs56149945), NR3C1 expression, and blood cortisol levels interacted to predict hippocampal volume, with higher NR3C1 expression being linked to smaller hippocampal volume (101). Additionally, FKBP5 expression in female mice is associated with reduced hippocampal neuron density (7). No significant associations were found between peripheral DNAm and serotonergic markers in postmortem brain samples (102). Overall, hippocampal development is shaped by experiences, likely through epigenetic mechanisms in the peripheral and central nervous systems at both the single-cell and population levels (103). Further study of the peripheral epigenome’s relationship with brain traits is needed to understand experience-driven hippocampal circuits that underlie cognitive and mental health.

Despite these important findings, several limitations should be noted. Saliva, buccal, and blood samples vary in correlation with brain DNAm (26, 27, 28), so using different peripheral sample types, such as blood or buccal cells, could yield different results. Adding a second peripheral sample type may improve interpretation. Our sample, primarily from a northeastern state with 83.1% White and 17.1% Hispanic/Latino individuals (104), is relatively diverse at 68.5% White and 21.0% Hispanic/Latino. However, broader racial diversity is limited, which may restrict generalizability. Prenatal nicotine exposure, known to impact DNAm and hippocampal volumes (105,106), was not controlled, potentially affecting associations and warranting future study. While no SLC6A4 SNP moderators of DNAm and hippocampal volume were detected, our small sample size limits these exploratory results. Larger cohorts are needed to replicate these findings and explore relationships across additional brain regions relevant to mental health. As an observational study, causality between brain structure and peripheral epigenetics should be inferred cautiously; here, predict refers only to associations, not causation. Lastly, it is known that serotonin influences spine density in the hippocampus (107,108), but such changes may not alter overall hippocampal volume (109) and thus would go undetected in this analysis.

Conclusions

Epigenetic processes like DNAm offer a promising molecular system for understanding the gene-environment-neurodevelopment interactions that influence mental health. This study is a valuable contribution, addressing gaps in psychiatric and neuroimaging epigenetics research (78), especially in pediatric or adolescent cohorts (62). Evidence increasingly shows that epigenetic mechanisms are crucial in shaping neurodevelopment, which influences behavior across the life span. Three key themes have emerged from animal and human studies: first, early environments interact with genetics to influence peripheral and central nervous system epigenetic markers; second, variations in these profiles shape typical and atypical neurodevelopment and mental health; and third, peripheral epigenetic markers may reflect brain changes in a spatially and temporally dependent way. Thus, further exploration of peripheral brain–epigenome relationships could help identify biological dysregulation and inform targeted, personalized treatments for maladaptive behavior.

Supplementary Material

Document S1

Tables S1 and S2