Integrative analysis of genomic and exposomic influences on youth mental health

Abstract

Background

Understanding complex influences on mental health problems in young people is needed to inform early prevention strategies. Both genetic and environmental factors are known to influence youth mental health, but a more comprehensive picture of their interplay, including wide‐ranging environmental exposures – that is, the exposome – is needed. We perform an integrative analysis of genomic and exposomic data in relation to internalizing and externalizing symptoms in a cohort of 4,314 unrelated youth from the Adolescent Brain and Cognitive Development (ABCD) Study.

Methods

Using novel GREML‐based approaches, we model the variance in internalizing and externalizing symptoms explained by additive and interactive influences from the genome (G) and modeled exposome (E) consisting of up to 133 variables at the family, peer, school, neighborhood, life event, and broader environmental levels, including genome‐by‐exposome (G × E) and exposome‐by‐exposome (E × E) effects.

Results

A best‐fitting integrative model with G, E, and G × E components explained 35% and 63% of variance in youth internalizing and externalizing symptoms, respectively. Youth in the top quintile of model‐predicted risk accounted for the majority of individuals with clinically elevated symptoms at follow‐up (60% for internalizing; 72% for externalizing). Of note, different domains of environmental exposures were most impactful for internalizing (life events) and externalizing (contextual including family, school, and peer‐level factors) symptoms. In addition, variance explained by G × E contributions was substantially larger for externalizing (33%) than internalizing (13%) symptoms.

Conclusions

Advanced statistical genetic methods in a longitudinal cohort of youth can be leveraged to address fundamental questions about the role of ‘nature and nurture’ in developmental psychopathology.

Introduction

Young people are at elevated risk of mental health problems, with epidemiological data suggesting about half of all mental disorders begin by mid‐adolescence (Kessler et al., ²⁰⁰⁷). Whether internalizing (e.g. depression/anxiety) or externalizing (e.g. problem behavior) in nature, early‐onset mental health problems are associated with poor lifelong outcomes including recurring psychiatric difficulties, suicidal behavior, and impaired functioning (Clayborne, Varin, & Colman, ²⁰¹⁹; Newman et al., ¹⁹⁹⁶; Woodward & Fergusson, ²⁰⁰¹). As such, understanding complex influences on mental health problems in young people is needed to inform strategies for preventing these problems and their downstream consequences.

Broadly speaking, it is known that both genetic and environmental factors play a role in influencing mental health among youth (Kwong et al., ²⁰¹⁹; Nikstat & Riemann, ²⁰²⁰). However, a complete picture remains obscured by the limited focus of prior studies. First, while psychiatric genetic research has benefited from moving to genome‐wide approaches (Sullivan & Kendler, ²⁰²¹), comprehensive analyses of environmental influences have been relatively sparse. Efforts to model environmental influences have typically captured one or a handful of exposures (e.g. childhood trauma) rather than a wider range of factors that could influence mental health (Guloksuz, van Os, & Rutten, ²⁰¹⁸). The exposome, defined as the totality of environmental exposures and their interactions (Wild, ²⁰¹²), is critical for understanding human health (Manrai et al., ²⁰¹⁷). Although our recent work has demonstrated the utility of exposure‐wide approaches in depression research (Choi et al., ²⁰²⁰), the cumulative exposome has remained challenging to model. Second, assessments of how environmental exposures interact with genome‐wide influences to shape mental health (Uher & Zwicker, ²⁰¹⁷) are limited, particularly in youth; existing genome‐wide environment interaction studies of depression, for example, have been conducted primarily in adult populations (e.g. Arnau‐Soler et al., ²⁰¹⁹; Coleman et al., ²⁰²⁰; Dunn et al., ²⁰¹⁶). If an interaction exists, this suggests that genetic influences on youth mental health vary by environmental factors and/or that environmental influences differ according to genetic profiles, pointing to more precise understanding of which individuals may be most susceptible or protected. Third, potential sources of bias in gene–environment studies, such as multicollinearity between environmental exposures which is common in psychosocial data and known to bias variance estimates (Zhou & Lee, ²⁰²¹), must be explicitly addressed.

Recent methods in statistical genetics can help overcome these gaps by allowing for the integrative modeling of genome‐wide (genomic) and exposure‐wide (exposomic) influences and their interactions, while accounting for potential estimation biases (Zhou & Lee, ²⁰²¹). Briefly, these methods rely on characterizing genome‐wide similarity among pairs of unrelated individuals against their phenotypic similarity and can now be extended to the estimation of exposure‐wide similarities (Zhou & Lee, ²⁰²¹). The Adolescent Brain and Cognitive Development (ABCD) Study – which includes genomic data, wide‐ranging environmental exposures, and rich longitudinal measurements of mental health in a cohort of more than 10,000 youth (Jernigan, Brown, & ABCD Consortium Coordinators, ²⁰¹⁸) – provides an important opportunity to conduct these comprehensive analyses. Here, we leveraged these data to further address fundamental questions about the role of ‘nature and nurture’ in developmental psychopathology. Using a novel linear mixed model approach, we sought to decompose the main and interactive effects of the genome and exposome on internalizing and externalizing symptoms in preadolescent youth. Advancing our understanding of the complex interplay of genes and environments in youth mental health may inform for whom and how best to focus preventative efforts.

Methods

Sample

The Adolescent Brain and Cognitive Development (ABCD) Study is a multisite longitudinal cohort study following over 10,000 adolescents beginning at the ages of 9–10. ABCD participants were recruited through schools across 21 sites in the United States using multistage probability sampling to capture a nationally representative sample (Garavan et al., ²⁰¹⁸). The study collected data from youth and parents across major domains including substance use, mental and physical health, neurocognition, culture and environment, and brain imaging (Karcher & Barch, ²⁰²¹). More information regarding ABCD study recruitment, design, and measures can be found elsewhere (Karcher & Barch, ²⁰²¹). Ethics approval for the parent study was obtained from all relevant institutional review boards, with written informed consent obtained from parents and verbal assent obtained from youth participants; secondary analyses of the publicly released, anonymized ABCD data were deemed by our local review board as not human subjects research. For this study, we examined the subsample with genomic data (see Appendix S1 for genomic QC and imputation details). Because estimates of whole‐genome effects can be confounded by population stratification, we focused on the largest European ancestry subgroup for primary demonstration (n = 5,369), retaining 6,834,371 variants with imputation INFO score ≥0.80 and minor allele frequency ≥1%. The top 10 genetic principal components (PCs) were used to adjust for residual stratification in subsequent analyses. We additionally retained one randomly selected individual from each related pair (pi‐hat > 0.20) resulting in a sample of genetically unrelated individuals. Our analytic sample consisted of 4,314 genetically unrelated youth with genomic data and baseline and 1‐year follow‐up symptom scores [47% female, mean age (SD; range) = 9.9 (0.6; 8.9–11.0) years] based on the ABCD 3.0 release. To leverage the diversity of the ABCD cohort, we also identified participants from non‐European groups with at least 1,000 unrelated individuals – of African (n = 1,125) and admixed American (n = 1,474) ancestry, respectively – to examine the consistency of our results.

Measures

Youth internalizing and externalizing symptoms

The parent‐completed Child Behavior Checklist (CBCL; Achenbach, ²⁰⁰¹) was used to index current youth internalizing and externalizing symptoms. The internalizing composite scale of the CBCL comprises the following subscale domains: Anxious/Depressed (e.g. cries a lot, feels worthless or inferior), Withdrawn/Depressed (e.g. very little he/she enjoys, underactive, sad, depressed), and Somatic Complaints (e.g. headaches, nausea, feels sick, stomachaches). The externalizing composite scale of the CBCL comprises the following subscale domains: Rule‐Breaking Behavior (e.g. lying or cheating, swearing or obscene language) and Aggressive Behavior (e.g. argues a lot, destroys his/her own things). For our main analyses, we used continuous internalizing (INT) and externalizing (EXT) T‐scores (typically scaled at mean 50 and SD 10) at the 1‐year follow‐up assessment. For sensitivity analyses, we also created a dichotomous indicator of clinically elevated symptoms at baseline, that is, T‐scores ≥70 (Pandolfi, Magyar, & Dill, ²⁰¹²).

Environmental exposures

Environmental exposures in our modeled exposome consisted of 133 survey‐based and/or derived variables (Tables S1 and S2) including proximal contextual variables (which may influence youth’s day‐to‐day experiences directly, for example, peer, family, and school factors; k = 36), broader contextual variables (which reflect broader environments surrounding youth, for example, neighborhood and geocoded environmental factors; k = 49), and life events variables (which include potential stressors, for example, injury, illness, loss, and family conflict; k = 48). All exposures were measured at baseline except life events, which were reported at 1‐year follow‐up but pertained to discrete events occurring prior to follow‐up, so were also included. We included variables reported by parents and/or youth where available, which may tap into different risk and protective experiences. Following data preprocessing (Appendix S2), we observed low missingness (0%–4%) among exposures and imputed these values for sample size consistency using the missForest package in R (Stekhoven & Bühlmann, ²⁰¹¹), a nonparametric method that efficiently handles mixed variable types and performs imputation by fitting iterative random forests (k = 100) on observed data to predict missing data.

Statistical analyses

Integrative analysis of genomic and exposomic data (IGE)

Using linear mixed models with genome‐based restricted maximum likelihood (GREML), we performed a comprehensive decomposition of genomic and exposomic influences on our two outcomes of interest: internalizing (INT) and externalizing (EXT) symptom scores at the 1‐year follow‐up. Scores were preadjusted (i.e. taking residuals of a regression model adjusted for the following variables) for youth age at baseline, sex, study site, and top 10 ancestry‐specific genetic PCs, then standardized to mean 0 and SD 1. To estimate the proportion of outcome variance attributable to additive and nonadditive (statistical) effects of the whole genome and modeled exposome – including interactions between exposome and genome, and between exposomic variables – we constructed a set of relationship matrices (Appendix S3 for matrix equations).

First, a relatedness matrix capturing genome‐wide variation (G) across unrelated individuals was generated in PLINK 1.9 (Chang et al., ²⁰¹⁵). Second, the exposome matrix (E) was generated using MTG2 v2.22 (Lee & van der Werf, ²⁰¹⁶) after transformation of the exposure data using principal components analysis (PCA). Specifically, we extracted the eigenvectors of the k × k covariance matrix of exposures (equivalent to the right singular vectors of the n × k exposure matrix) via eigen decomposition. We then multiplied the exposure matrix by the eigenvectors to produce a column‐orthogonal, and subsequently column‐standardized, matrix for fitting the linear mixed models. For greater detail, see Zhou and Lee (2021). For the E matrix, we used k = 133 environmental factors that included proximal contextual variables, broader contextual variables, and prior life events described above. Third, the genome‐by‐exposome interactions matrix (G × E) was computed as the Hadamard product of the transformed exposome matrix and genomic relatedness matrix. Fourth, an exposome‐by‐exposome interactions matrix (E × E) was generated from the pairwise products between all environmental factors, again pretransformed to account for potential multicollinearity. Coding procedures, including steps to pretransform exposures and generate relationship matrices, are outlined in detail in the MTG2 manual (https://sites.google.com/view/s‐hong‐lee‐homepage/mtg2).

Variance in INT or EXT symptoms explained by these components was estimated in the following nested models: (a) G only; (b) G and E; (c) G, E, and G × E; and (d) G, E, G × E, and E × E. Each model was statistically evaluated against the previous model using likelihood ratio tests, with a significant result (p < .0083 = .05/6 based on three model comparisons for two outcomes), indicating model fit improvement due to the additional variance component. The model after which no significant improvement was detected was considered the best‐fitting model. Variance component estimates whose 95% confidence intervals did not contain zero were interpreted as significant. We then used a fivefold cross‐validation approach to index the predictive accuracy of the best‐fitting model for each outcome by generating model‐predicted values in 80% of the sample (training) and calculating Pearson correlations with observed outcome values in the remaining 20% holdout sample (testing), averaging these correlations across five nonoverlapping random splits, as in prior work (Zhou & Lee, ²⁰²¹). For further interpretation, we divided individuals in each testing set based on quintiles of model‐predicted values (based on best linear unbiased predictions summed across all random effect components) and, using the metafor R package, meta‐analyzed the beta increase in standardized symptom scores associated with membership in each prediction quintile compared to the bottom reference quintile. We also calculated the average risk concentration (i.e. proportion of all youth with clinically elevated symptoms at follow‐up; n = 154 for INT and n = 81 for EXT) represented across model‐predicted quintiles. We then performed sensitivity analyses to evaluate the best‐fitting model and its variance components for consistency: (i) after removing individuals with clinically elevated symptoms at baseline (n = 140 for INT and n = 75 for EXT), (ii) when adjusting for parental educational attainment, (iii) including perinatal exposures, or (iv) without pretransforming exposures for independence. We also assessed potential differences by reporter source by modeling the exposome component based on child‐reported (k = 61) versus parent‐reported (k = 39) exposures. Though underpowered, we also evaluated the best‐fitting model in the admixed American and African ancestry groups.

Finally, where a significant interaction (e.g. G × E) variance component was observed, we further decomposed the E matrix into major subdomains – proximal contextual variables (k = 36), broader contextual variables (k = 49), and life events variables (k = 48) – to separately examine the influence of each subdomain and assess its potential role in the interaction. Lastly, we fitted the best‐fitting model including all subdomain exposome (and genome‐exposome) matrices simultaneously.

Results

Variance decomposition results for INT and EXT, estimating joint contributions of genomic and exposomic influences on mental health symptoms in ABCD participants, are reported in Table 1. First, when examined alone (G model), the G variance component (${\sigma }_{\mathrm{G}}^2$) was modest and nonsignificant for INT symptoms (6%), but sizeable and significant for EXT symptoms (19%). Second, including additive effects of the exposome (G and E model) improved the genomics‐only model. Here, ${\sigma }_{\mathrm{G}}^2$ remained modest for INT (6%) but was markedly attenuated for EXT (8%), while the E variance component (${\sigma }_{\mathrm{E}}^2$) was estimated at 16% for INT and 23% for EXT. Third, including interactive effects between the exposome and genome (G, E, and G × E model) showed further improvement over the previous model. Here, ${\sigma }_{\mathrm{G}}^2$ and ${\sigma }_{\mathrm{E}}^2$ remained virtually unchanged in magnitude, while the G × E variance component (${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$) was estimated to be 13% for INT and 33% for EXT, indicating sizeable unique variance explained by genome‐exposome interactions. Fourth, including nonadditive effects of the exposome (G, E, G × E, and E × E model) did not result in significant improvements. Here, the E × E variance component (${\sigma }_{\mathrm{E}\times \mathrm{E}}^2$) accounted for <1% of phenotypic variance in INT and EXT symptoms, while other component estimates remained unchanged. Thus, the more parsimonious G, E, and G × E model was considered to be the best‐fitting model for both outcomes. The proportion of phenotypic variance in INT and EXT symptoms explained by each component in the best‐fitting model is visualized in Figure 1.

Table 1.

Genomic and exposomic variance components for youth internalizing and externalizing symptoms

Model	G only	G and E	G, E, and G × E	G, E, G × E, and E × E
INT
${\sigma }_{\mathrm{G}}^2$	0.058 (−0.092, 0.208)	0.06 (−0.069, 0.19)	0.063 (−0.067, 0.191)	0.063 (−0.066, 0.192)
${\sigma }_{\mathrm{E}}^2$		0.161 (0.116, 0.206) a	0.161 (0.115, 0.205) a	0.161 (0.115, 0.205) a
${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$			0.129 (0.042, 0.215) a	0.126 (0.037, 0.215) a
${\sigma }_{\mathrm{E}\times \mathrm{E}}^2$				−0.002 (−0.024, 0.019)
LKH	−2,160.3820	−1,914.1290, p = 4.06e‐109 b	−1,909.7914, p = .0032 b	−1,909.7730, p = .848
EXT
${\sigma }_{\mathrm{G}}^2$	0.192 (0.036, 0.347) a	0.084 (−0.036, 0.203)	0.074 (−0.042, 0.189)	0.072 (−0.044, 0.187)
${\sigma }_{\mathrm{E}}^2$		0.233 (0.171, 0.294) a	0.226 (0.165, 0.286) a	0.226 (0.165, 0.286) a
${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$			0.327 (0.239, 0.414) a	0.334 (0.242, 0.424) a
${\sigma }_{\mathrm{E}\times \mathrm{E}}^2$				0.005 (−0.014, 0.024)
LKH	−2,157.5381	−1,745.7653, p = 4.10e‐181 b	−1,715.3967, p = 6.52e‐15 b	−1,715.2634, p = .606

σ ² = variance components reflect proportions of total variance estimated by the model; residual (e) components not shown.

^a Nonzero overlapping 95% confidence interval.

^b Significant model likelihood ratio test compared to previous nested model.

Figure 1.

Estimated total proportion of variance in INT and EXT symptoms explained by G, E, and G × E components. e, residual variance; E, modeled exposome effects based on 133 exposure variables; G, genomic effects; G × E, genome‐by‐exposome effects [Colour figure can be viewed at wileyonlinelibrary.com]

In fivefold cross‐validation, we found that predicted values by the best model (G, E, and G × E) were significantly correlated with observed values of INT [fivefold average r = .36 (range = .32–.43), p < 2.2e‐16] and EXT [fivefold average r = .45 (range = .40–.51), p < 2.2e‐16] in the holdout sets. Individuals in the top quintile of model‐predicted values had symptom scores at least 1 SD or higher compared to those in the bottom quintile (B = 1.00 and 1.25 for INT and EXT, respectively; Figure 2A) in the holdout sets, translating to roughly 10–12 points higher on their CBCL T‐score. The upper quintiles of model‐predicted values also accounted for the large majority of youth with clinically elevated symptoms at follow‐up (81% for INT and 91% for EXT in the top two quintiles; 60% and 72% in the top quintile alone; Figure 2B).

Figure 2.

Integrative (G, E, and G × E) model predictions of elevated internalizing and externalizing symptoms at follow‐up. (A) Beta values associated with each model‐predicted quintile (meta‐analyzed across the fivefold holdout splits) are shown with 95% CI; individuals in quintile 1 of model‐predicted values are considered as the reference group. Dashed line shows 1 SD increase in symptoms. (B) Risk concentrations (i.e. proportions of all youth with clinically elevated symptoms for each follow‐up outcome) are shown in the shaded bar below for each quintile [Colour figure can be viewed at wileyonlinelibrary.com]

In sensitivity analyses, ${\sigma }_{\mathrm{E}}^2$ and ${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$ were minimally attenuated when further adjusting for parental educational attainment (Table S3), did not increase substantially when including perinatal exposures (Table S4), and remained significant albeit slightly attenuated, by about 2%–4%, after excluding participants with elevated baseline symptoms (Table S5). When exposures were not initially transformed for independence, ${\sigma }_{\mathrm{E}}^2$ and ${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$ were reduced but remained significant (Table S6). When separating child‐ and parent‐reported exposures (Table S7), significant ${\sigma }_{\mathrm{E}}^2$ and ${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$ were observed for both outcomes regardless of child‐ or parent‐reported exposures, with slightly higher ${\sigma }_{\mathrm{E}}^2$ based on parent‐reported exposures and slightly higher ${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$ based on child‐reported exposures. For the admixed American and African ancestry groups (Table S8), ${\sigma }_{\mathrm{E}}^2$ for both outcomes was similar to the European ancestry group; ${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$ was qualitatively larger in the African ancestry group across both outcomes; and ${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$ was significant for EXT but not INT in the admixed American ancestry group.

When decomposing the full E matrix and re‐estimating the G, E, and G × E model where E comprised only variables pertaining to each subdomain – proximal contextual factors, broader contextual factors, or life event factors – a significant ${\sigma }_{\mathrm{E}}^2$ persisted for all three subdomains (Table 2). Comparing across the three subdomains, the largest ${\sigma }_{\mathrm{E}}^2$ was observed when E comprised life events for INT (11%) and when E comprised proximal contextual variables for EXT (16%). For both outcomes, ${\sigma }_{\mathrm{E}}^2$ was smallest in magnitude when E comprised broader contextual variables (3% for INT; 5% for EXT), though still nonzero. Notably, ${\sigma }_{\mathrm{G}}^2$ was substantially reduced for EXT when jointly modeled with ${\sigma }_{\mathrm{E}}^2$ based on proximal contextual or life events factors, but remained relatively intact (i.e. significant) when jointly modeled with ${\sigma }_{\mathrm{E}}^2$ based on broader contextual variables. On the other hand, ${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$ was only significant when E comprised either proximal contextual variables or life event variables, but not when E comprised broader contextual variables – with the largest ${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$ based on E comprising proximal contextual variables (14% and 26% for INT and EXT, respectively). These results were consistent with simultaneous models (Tables S9 and S10) where, even considered together, all three exposome subdomains contributed significant ${\sigma }_{\mathrm{E}}^2$ across both outcomes, while proximal contextual and life event subdomains both contributed significant ${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$, but the broad contextual subdomain did not.

Table 2.

Decomposition of best‐model components based on exposomic subdomains

	Proximal contextual (k = 36)	Broader contextual (k = 49)	Life events (k = 48)
INT
${\sigma }_{\mathrm{G}}^2$	0.053 (−0.085, 0.191)	0.060 (−0.087, 0.206) b	0.055 (−0.08, 0.189)
${\sigma }_{\mathrm{E}}^2$	0.079 (0.039, 0.119) a	0.029 (0.013, 0.045) a	0.108 (0.06, 0.155) a , b
${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$	0.129 (0.058, 0.198) a , b	−0.012 (−0.04, 0.016)	0.119 (0.061, 0.176) a
EXT
${\sigma }_{\mathrm{G}}^2$	0.079 (−0.044, 0.201)	0.167 (0.019, 0.316) a , b	0.134 (0, 0.268)
${\sigma }_{\mathrm{E}}^2$	0.163 (0.084, 0.242) a , b	0.045 (0.023, 0.067) a	0.112 (0.062, 0.161) a
${\sigma }_{\mathrm{G}\times \mathrm{E}}^2$	0.280 (0.21, 0.35) a , b	0.038 (−0.026, 0.102)	0.199 (0.138, 0.258) a

σ ² = variance components reflect proportions of total variance estimated by the model; residual (e) components not shown.

^a Nonzero overlapping 95% confidence interval.

^b Largest variance explained by that component (row‐wise) across the three exposomic subdomains: Proximal contextual = E comprising 36 friends/caregiver/family/school variables; Broader contextual = E comprising 49 neighborhood/area‐level variables; Life events = E comprising 48 life event variables.

Discussion

In this study, we applied advanced statistical genetics methods in a longitudinal cohort of preadolescent youth to address fundamental questions about the role of ‘nature and nurture’ in developmental psychopathology. Because of our ability to explicitly model joint genomic and exposomic effects as well as their potential interactions, this represents to our knowledge the most comprehensive decomposition to date of genetic and environmental influences on mental health in youth using molecular genetic data.

Key findings

Several notable findings emerged. First, an integrative model including additive and interactive influences of the genome and exposome explained over 30% and 60% of variance in youth internalizing and externalizing symptoms at follow‐up, respectively. By contrast, a single variable with known impacts on externalizing behavior – parental separation/divorce (Garriga & Pennoni, ²⁰²⁰), which was associated with externalizing symptoms in our sample (p < 1e‐10 in adjusted linear regression) – accounted for only 1.1% of its variance. Moreover, joint models of polygenic risk and a small selection of environmental exposures in a recent study of 1,154 preadolescent youth in the Netherlands accounted for no more than 11%–12% variance of internalizing and externalizing symptoms (Ensink et al., ²⁰²⁰), highlighting the enhanced explanatory power of our integrative genomic and exposomic modeling. In fivefold cross‐validation, youth in the top quintile of predicted values from the integrative model had symptom T‐scores that were on average 10–12 points (at least 1 SD) higher than those in the bottom quintile. For context, an increase of 15 points above the mean T‐score reflects ‘borderline’ clinical risk (Achenbach, ²⁰⁰¹). Moreover, among all youth with clinically elevated internalizing and externalizing symptoms at follow‐up, the majority (80%–90%) were captured in the upper distribution of model‐predicted values.

Second, exposomic influences generally overshadowed genomic contributions to mental health symptoms in preadolescent youth, perhaps unsurprisingly. The additive modeled exposome explained a substantial proportion of variation in observed symptoms (16% for internalizing and 23% for externalizing). Among exposomic subdomains, life events such as experiences of injury, illness, or loss appeared most influential for internalizing symptoms in youth, substantiating known relationships between stressful experiences and depression (Hammen, ²⁰⁰⁵; Kessler, ¹⁹⁹⁷; Mazure, ¹⁹⁹⁸), while proximal contextual variables including family, school, and peer‐level factors were most influential for externalizing symptoms, suggesting that everyday contexts where a child is embedded have some of the most powerful influences on problem behavior (Bank, Burraston, & Snyder, ²⁰⁰⁴; Tomé, Matos, Simões, Diniz, & Camacho, ²⁰¹²). Broader contextual factors including neighborhood crime and area‐level deprivation, while non‐negligible in their influences, contributed less to overall symptom variance than proximal contextual variables or life events.

We initially found that genome‐wide influences explained more than twice the variance in externalizing symptoms (19%) compared to internalizing symptoms (6%) at this age. This aligns with twin‐based literature, indicating externalizing symptoms are more heritable than internalizing symptoms in early adolescence (Bergen, Gardner, & Kendler, ²⁰⁰⁷; Nikstat & Riemann, ²⁰²⁰) – potentially due to measurement, as externalizing behaviors are more easily observed and potentially captured with reduced noise (Bartels et al., ²⁰⁰⁴), or life course timing, since depression and other internalizing conditions may onset later in adolescence and young adulthood (Nivard et al., ²⁰¹⁷). Unlike twin studies that extrapolate genetic and environmental influences based on inherited similarities/differences, we examined common genomic variation in unrelated individuals. While not significant, our G estimate for internalizing symptoms is similar to reported SNP‐based heritability of self‐reported internalizing symptoms (6%) from a recent meta‐analytic Genome‐Wide Association Study (GWAS) of 20 childhood and adolescent cohorts spanning a broader age range (Jami et al., ²⁰²¹).

When modeled alongside the exposome, genomic effects on internalizing symptoms remained unchanged in magnitude, suggesting relative independence from exposomic effects, whereas genomic effects on externalizing symptoms were markedly attenuated. Our findings suggest that the modeled exposome may account for more than half of the estimated genome‐wide effects on externalizing symptoms. It could be that genomic effects on externalizing symptoms are mediated through environmental pathways; for example, genome‐wide variation may increase the likelihood of certain exposures (e.g. negative life events, Johnson, Rhee, Whisman, Corley, & Hewitt, ²⁰¹³), which then influence externalizing risk. Alternatively, genetic factors that underlie endophenotypes of externalizing behavior (e.g. impulsivity) may also shape environmental factors (e.g. family conflict), reflecting gene–environment correlation (Lau & Eley, ²⁰⁰⁸), though our results remained largely consistent when excluding youth with clinically elevated externalizing symptoms at baseline. Genome‐wide effects were most attenuated when the modeled exposome consisted of proximal contextual or life events variables, indicating these factors may be particularly intertwined with the genetic basis of externalizing behavior. If so, genetic risk for youth problem behavior could potentially be modified by interventions targeting these environmental factors.

Third, genome–exposome interactions appear to contribute substantially to youth mental health symptoms. Modeling these interactions explained an additional 13% of variation in internalizing and an additional 33% in externalizing symptoms. This suggests a key role for ‘nurture’ since exposomic influences on youth mental health were not only more profound on their own, but appear to also modulate genomic influences as well. Interestingly, genome‐wide variation appeared to interact mostly with proximal contextual factors or life events, rather than broader contextual variables, in shaping youth symptoms. On the other hand, exposome‐by‐exposome interaction effects were modest in boosting explanatory power, suggesting the environmental factors we considered in this sample are more likely to operate additively. However, accounting for their potential intercorrelations appeared to increase exposomic signal. In sensitivity analyses where exposures were not initially transformed for independence, the estimated exposome effect was qualitatively reduced albeit still significant. One potential interpretation is that we may generally underestimate exposomic effects due to multicollinearity. Simulation work has shown that correlation structure among exposures can produce biased estimates (Zhou & Lee, ²⁰²¹) as the effective number of exposures becomes smaller than the actual number of modeled exposures.

Strengths and limitations

This study offers numerous strengths, including the explicit modeling of an ‘exposome’ based on 133 different variables (at family, peer, school, neighborhood, life event, and broader environmental levels), and longitudinal design with temporal separation of modeled exposures from outcomes at follow‐up. We were able to further dissect the exposome into conceptual subdomains to examine relative contributions of multilevel factors from additive and interactive perspectives. Furthermore, we modeled genome‐wide similarity and differences across individuals, rather than relying on polygenic scores that largely capture additive effects from GWAS for specific traits only and to date explain limited phenotypic variance in psychiatric outcomes (Murray et al., ²⁰²¹). For example, a recent study from the ABCD cohort reported even the most predictive polygenic scores (e.g. ADHD PRS) explained no more than 1.4% of variance in CBCL‐related outcomes (Waszczuk et al., ²⁰²¹).

Several limitations, however, should be noted. First, it is worth considering that despite modeling a set of genome, exposome, and genome‐by‐exposome influences, a notable proportion of variance in youth mental health symptoms remained ‘missing’, or unexplained. This might be explained by the fact that our modeled exposome, while wide‐ranging in scope, was limited to assessed variables rather than all possible environmental exposures that could shape youth mental health, including those occurring earlier in life, though we found including several common perinatal exposures contributed minimally to additional variance. We also did not examine the role of rare or X‐linked genomic variation. It may be that no set of genomic and/or exposomic factors is deterministic of youth mental health symptoms, raising the importance of modeling idiographic variation. Second, this was a study of mental health symptoms in preadolescent youth, whereas the heritability – and epidemiological risk – of internalizing and externalizing phenotypes tends to increase over time in young people (Bergen et al., ²⁰⁰⁷). However, early identification of at‐risk individuals is arguably timely for life course prevention. Third, despite leveraging the largest genetic ancestry group for demonstration, our sample size was relatively modest for detecting significant genome‐wide effects for complex psychiatric traits, though the magnitudes of heritability estimates were not dissimilar from other recent studies. Fourth, there may be reverse influences of internalizing/externalizing risk on the modeled exposome; directionality is not established, although we sought to temporally separate outcome measures from the environmental exposures at baseline or prior to follow‐up. Finally, symptom outcomes were parent‐reported, which may differ from other informants such as teachers or youth themselves. Shared method variance where outcome and exposures are reported by the same informant (Xerxa et al., ²⁰²¹) may upward bias the exposome estimates, although we found that exposome and genome‐exposome influences persisted even when relying solely on child‐reported exposures. Because some exposures were only reported by child versus parent (e.g. school vs. neighborhood factors), we cannot fully untangle shared method bias from substantive differences in assessed exposures; this may require studies where environmental exposures have been consistently measured across multiple informants. For generalizability, our further replications in various age brackets of youth from this and other large‐scale developmental cohorts, with different modes of measurement and across diverse settings, are warranted and currently ongoing. Certain environmental exposures (e.g. neighborhood contextual factors) may influence mental health more substantially as older youth develop greater awareness of contexts outside of home and school and may suggest the need for dynamic prevention approaches that shift in their targets over time.

Conclusion

In conclusion, the relative contributions and interplay between genes and environment has been a central focus of developmental psychopathology. Applying recent advances in statistical genetic modeling to a longitudinal cohort comprising genomic, exposomic, and clinical data, we demonstrate the influence of multilevel contexts and their interaction with genome‐wide variation on emotional and behavioral problems in young people. Identifying exposomic components of mental health risk may inform the design of prevention strategies.

Supporting information

Appendix S1. Genomic data processing.

Appendix S2. Phenotypic data pre‐processing.

Appendix S3. Relationship matrix specifications.

Table S1. Assessment of exposures in the study cohort.

Table S2. More information on exposures.

Table S3. Estimated variance components for G, E, G × E model when further pre‐adjusting outcomes for parental educational attainment.

Table S4. Estimated variance components for G, E, G × E model when including perinatal exposures.

Table S5. Estimated variance components for G, E, G × E model when excluding individuals with clinically elevated baseline symptoms.

Table S6. Estimated variance components for G, E, G × E model where E components were not initially PCA‐transformed.

Table S7. Estimated variance components for G, E, G × E model based on only parent‐ versus child‐reported exposures.

Table S8. Estimated variance components for G, E, G × E model in the African and admixed American ancestry groups.

Table S9. Estimated variance components for G, E, G × E model of INT where multiple E subdomains were simultaneously modeled.

Table S10. Estimated variance components for G, E, G × E model of EXT where multiple E subdomains were simultaneously modeled.