Relationships Between Polygenic Scores for Psychopathology and Observed Psychopathology Are Mediated by Cognitive Control and Reward Sensitivity Pathways: Insights From the ABCD Study

Abstract

Background

Psychiatric disorders affect approximately 20% of adolescents in the United States and are both heritable and highly polygenic. Polygenic scores (PGSs) aggregate genetic risk across the genome and in doing so potentially capture multiple, distinct behavioral pathways. Among these, cognitive control and reward sensitivity are key constructs that have been found to be independently and jointly related to adolescent psychopathology.

Methods

In this study, we used latent variables representing cognitive control and reward sensitivity (measured at baseline) as mediators between PGSs for internalizing, neurodevelopmental, and addiction-related psychopathology and observed psychopathology measured 2 years later in the ABCD (Adolescent Brain Cognitive Development) Study (N = 11,875; ages 9–12 years; 48% female, 52% male assigned at birth).

Results

Our findings show that most of the associations between PGSs and later psychopathology were accounted for by indirect effects through cognitive control and reward sensitivity; all of the associations between the addiction PGS and externalizing were accounted for by the indirect effects, but less of the internalizing PGS to internalizing pathway was accounted for by control and reward sensitivity. Multiple cognitive control factors significantly mediated the effects of PGSs on both internalizing and externalizing outcomes (indirect effects = 0.004–0.074). Reward sensitivity, specifically sensation seeking, uniquely mediated the association between the addiction PGS and later psychopathology, suggesting that addiction-related genetic risk may operate through pathways related to heightened reward responsiveness.

Conclusions

Overall, these results indicate that PGSs for internalizing, neurodevelopmental, and addiction-related traits index multiple distinct behavioral mechanisms with specific relevance to adolescent psychopathology risk.

Plain Language Statement

About 1 in 5 U.S. adolescents experience a psychiatric disorder, many of which have genetic underpinnings. In this study, we examined how genetic risk for internalizing, neurodevelopmental, and addiction-related traits influence later mental health through 2 key behaviors, cognitive control (the ability to manage thoughts and actions) and reward sensitivity (how strongly someone seeks or responds to rewards). Using data from nearly 12,000 youth, we found that most genetic effects on later mental health operated indirectly through these behaviors. In particular, sensation seeking linked addiction-related genetic risk to later psychopathology.

Understanding the etiology of psychiatric illnesses during adolescence is critical, as 1 in 7 youth globally (ages 10–19 years) experience a mental health disorder (1). Over the past decade, research investigating the genetic underpinnings of psychiatric disease (i.e., genome-wide association studies [GWASs]) has accelerated, particularly in adult samples (2,3). Findings from these genomic studies suggest both high polygenicity—where complex phenotypes are influenced by many genetic variants, each of which has a very small effect (4)—and widespread pleiotropy—where many of the same genetic effects are implicated across multiple disorders (4, 5, 6, 7, 8). However, it is unclear what biological and behavioral factors these genetic influences reflect. Reward sensitivity and cognitive control are behavioral constructs relevant to adolescent and adult psychopathology (9,10) that may mediate some of the shared genetic risk of psychiatric disorders. Here, we evaluated the relationships of genetic risk for psychopathology to reward sensitivity and cognitive control in the ABCD (Adolescent Brain Cognitive Development) Study to clarify the pathways between genetic liability for a disorder and symptoms during adolescence.

Because many complex traits, including psychiatric disorders, exhibit a highly polygenic structure, genetic risk for these disorders is not well-quantified with individual genetic variants. Rather, the effect sizes of risk alleles (obtained from a GWAS) can be summed across the entire genome to quantify an individual’s genetic risk with a polygenic score (PGS). PGSs have shown utility in predicting the emergence and severity of a given disorder (11, 12, 13). Although PGSs are often developed from GWASs of psychiatric traits measured in adulthood, they have been efficacious in predicting psychopathology in adolescent samples (14,15). This predictiveness suggests common genetic influences across adolescence and adulthood, which may be related to risk factors present in adolescence.

However, it is unknown what biological and behavioral factors these scores reflect. Given that a PGS captures genetic liability across the genome, it likely encapsulates multiple distinguishable biological pathways that collectively confer risk for psychopathology. Reward sensitivity and cognitive control are genetically influenced (16, 17, 18) behavioral constructs relevant to adolescent and adult psychopathology. Thus, we examined whether reward sensitivity and cognitive control mediate some of the shared genetic risk identified in GWASs of psychiatric disorders. Mediation models between PGSs for psychopathology-related phenotypes and observed psychopathology have been useful in past work (19).

Adolescent behavior is often conceptualized through models emphasizing the interplay between reward sensitivity and cognitive control, such as the maturational imbalance and dual-systems frameworks. The maturational imbalance model posits that reward-related brain regions mature earlier than those governing cognitive control (20), while dual-systems models describe risk-taking as arising from competition between bottom-up reward and top-down control processes (9). The importance of these constructs is further reflected in the National Institute of Mental Health’s Research Domain Criteria, which identifies reward responsiveness and cognitive control as key dimensions of psychopathology across multiple levels of analysis (21).

Our previous work (22) has extended these developmental psychopathology paradigms to characterize how reward sensitivity and cognitive control, measured by multiple raters (youths and parents) and multiple methods (tasks and surveys), are related to adolescent psychopathology in the ABCD Study, a national longitudinal study of more than 11,000 youth in the United States. Figure 1 shows this model, which suggested that 1) task-based cognitive control was separate from survey-based cognitive control (loading onto separate, moderately correlated factors); 2) multiple cognitive control factors, including task-based and survey-based factors, were transdiagnostic; for example, activation control was related to both internalizing and externalizing (Table S1); and 3) reward sensitivity and cognitive control factors were independently predictive of multiple forms of psychopathology (Table S1). This work confirmed that multiple dimensions of reward sensitivity and cognitive control are relevant to adolescent psychopathology.

Figure 1.

Schematic of mediation model and associated factors. Ovals indicate latent variables. Rectangles indicate observed items. Text boxes represent indicator items, but the figure does not show all individual items for simplicity. For exploratory structural equation model (ESEM) indicators, text boxes represent subscales that individual survey items originated from but do not represent sum scores or aggregate subscale scores. While we found 7 ESEM cognitive control/reward sensitivity factors, we did not include a latent variable for each one to further reduce complexity in the figure; however, there were 7 correlated ESEM factors. For the psychopathology factors, the text boxes represent individual diagnoses or subscale sum scores. Youth/parent survey factors are orthogonal method factors with equated individual survey item loadings. Youth/parent/teacher symptom factors are orthogonal method factors with equated individual symptom item loadings. ACT, activation control factor; ADHD, attention-deficit/hyperactivity disorder; BAS, Behavioral Activation Scale factor; EATQ, Early Adolescent Temperament Questionnaire; EF, executive function task factor; ESEM, exploratory structural equation model; INH, inhibitory control factor; OCD, obsessive-compulsive disorder; P-CBCL, parent-reported Child Behavior Checklist; P-KSADS, parent-reported Kiddie Schedule of Affective Disorders and Schizophrenia; PRE, lack of premeditation factor; PU, positive urgency factor; SS, sensation seeking factor; T-BPM, teacher-rated Brief Problem Monitor; UPPS, Urgency, Perseverance, Premeditation, and Sensation Seeking; Y-BPM, youth-rated Brief Problem Monitor; Y-KSADS, youth-rated Kiddie Schedule of Affective Disorders and Schizophrenia.

In this study, we 1) quantify the extent to which PGSs for adult psychopathology factors are predictive of psychopathology in an adolescent sample and 2) evaluate the extent to which these PGSs are capturing individual differences in cognitive control and reward sensitivity in adolescence. Specifically, we examine the extent to which the associations between PGSs and adolescent psychopathology are mediated by cognitive control and reward sensitivity factors.

Due to the pleiotropic nature of psychiatric disorders (23, 24, 25, 26, 27), we used GWAS summary statistics derived from multivariate genomic structural equation models (genomic SEMs) (14,28). We calculated PGSs for an addiction risk factor (28) and neurodevelopmental and internalizing psychopathology factors (29), consistent with our phenotypic model, which focuses on psychopathology factors rather than individual disorders. This approach reduces the number of statistical tests and recognizes the pleiotropic nature of genetic influences on psychiatric disorders (7).

We build on prior cross-sectional work (10,30,31) by leveraging the longitudinal nature of the ABCD Study and relating our predictor variables measured at baseline to psychopathology outcomes measured 2 years later. We are interested in whether the genetic liability for later psychopathology can be partially captured by behavioral risk factors that manifest earlier in adolescence. Utilizing longitudinal data also helps to guard against inflated relationships with psychopathology due to time-specific effects.

Methods and Materials

Participants

We included 11,875 participants who completed the baseline (mean age = 9.92 years, SD = 0.63) and year 2 follow-up (mean age = 12.03 years, SD = 0.67) assessments from the ABCD Study. Primary analyses included a subset of participants (n = 6667; sex = 48% female, 52% male) whose ancestry was estimated to be similar to European ancestry based on genomic principal components. The full ABCD sample was recruited to be demographically diverse and generally match the American Community Survey national estimates for demographic characteristics (32). Human subjects research was reviewed and approved by a central institutional review board and complies with the World Medical Association Declaration of Helsinki and American Psychological Association ethical standards. Written informed consent was provided by legal guardians. Data were sourced from the ABCD 5.1 release (https://doi.org/10.15154/z563-zd24).

Measures

Reward Sensitivity and Cognitive Control Measures

The phenotypic measures we used have been described in previous work and are shown in Tables S1 and S2, which detail the measures and factor structure of the reward sensitivity, cognitive control, and psychopathology measures. We chose behavioral tasks and survey items to assess reward sensitivity and cognitive control based on face validity and previous literature regardless of the original subscale. Items came from the Behavioral Activation Scale (BAS) (33); the Urgency, Perseverance, Premeditation, and Sensation-Seeking scale (34); and the Early Adolescent Temperament Questionnaire (35). We preregistered these item selections (https://doi.org/10.17605/OSF.IO/JKNR5). Overall surveys have all previously shown good reliability in adolescent samples (36,37).

Reward sensitivity tasks were delay discounting and game of dice. The following cognitive control tasks were drawn from prior work evaluating an executive function model (35): flanker inhibitory control (flanker), stop signal task, emotional n-back, list sorting working memory, and dimensional change card sort (card sort). Task methods and reliabilities are described in prior work (37,38) and shown in Table S3.

Psychopathology Measures

We used psychopathology measures from the year 2 follow-up to test whether the reward sensitivity and cognitive control factors measured at baseline were associated with psychopathology measured 2 years later. Because some individuals were missing data at the year 2 follow-up, we confirmed that there were not statistically significant differences between those missing or not missing data at year 2 based on psychopathology status (Table S4).

Psychopathology outcomes were described in detail in prior work (22). Youth self-report was measured with the Kiddie Schedule of Affective Disorders and Schizophrenia for DSM-5 (KSADS-5) and the Brief Problem Monitor (BPM) (39). A caregiver completed the Child Behavior Checklist (CBCL) (38) and KSADS about the child. The child’s primary school teacher completed the BPM-Teacher form (39).

The diagnostic classes utilized for this analysis and their corresponding endorsement rates are provided in Tables S5 to S7. A full list of the KSADS-5 disorders assessed and how disorder classes were generated (40) is provided Table S5. Descriptive statistics for the psychopathology measures are provided in Tables S6 to S7.

PGS Calculation

We used PRS-CS (polygenic risk score-continuous shrinkage) (41) and PLINK version 2.0 (42) to compute 3 PGSs (addiction, internalizing, and neurodevelopment). PRS-CS is a Bayesian polygenic prediction method that uses a continuous shrinkage prior to infer posterior single-nucleotide polymorphism (SNP) weights. Here, PRS-CS was used to adjust the weights for the correlated SNPs, with the 1000 Genomes Project used as reference panel for linkage disequilibrium (LD). The PLINK –score command was used to calculate a PRS for each individual by summing all variants weighted by the inferred posterior effect size.

Summary statistics for the addiction PGS were derived from a multivariate GWAS conducted using genomic SEM (14), incorporating GWASs of problematic alcohol use (λ = 0.58) (14, 43), problematic tobacco use (λ = 0.36) (14, 44), cannabis use disorder (λ = 0.93) (14, 45), and opioid use disorder (λ = 0.83) (14, 46) (Table S12). The neurodevelopmental and internalizing PGSs were similarly derived from multivariate GWASs (29). The neurodevelopmental factor included Tourette syndrome (λ = 0.14) (29, 47), attention-deficit/hyperactivity disorder (λ = 0.74) (29, 48), autism spectrum disorder (λ = 0.49) (29,49), problematic alcohol use (λ = 0.21) (29,43), and posttraumatic stress disorder (λ = 0.41) (29,50). The internalizing factor included major depressive disorder (λ = 0.92) (29,51), generalized anxiety disorder (λ = 0.73) (29,52), posttraumatic stress disorder (λ = 0.55) (29,50), and problematic alcohol use (λ = 0.15) (29,43) (Table S12). We did not include the externalizing PGS (53,54) because several of its component phenotypes overlapped with our mediator constructs. Specifically, the primary phenotype, general risk tolerance, shared conceptual overlap with sensation seeking, posing a risk of inflated estimates due to phenotypic confounding. Follow-up analyses using the externalizing PGS are reported in Tables S8 and S9.

We focused our results on PGSs for participants with genomic ancestry similar to Europeans because the discovery GWASs were restricted to individuals with European ancestry, and PGSs show limited portability across ancestries (55, 56, 57). Additional analyses on other ancestral groups (Table S10) showed inconsistent patterns, warranting cautious interpretation due to these portability issues and small sample sizes. Details on GWAS calculations, quality control pipeline, and ancestry estimation are in the Supplemental Methods.

Statistical Analyses

Statistical procedures for modeling reward sensitivity, cognitive control, and psychopathology outcomes followed previously established methods (22). A 7-factor exploratory SEM (ESEM) was selected after evaluating models with 1 to 9 factors. Psychopathology outcomes were modeled as confirmatory factor analysis factors within the ESEM framework (36,58). To account for shared rater variance, orthogonal rater factors were included for both survey-based (ESEM) and symptom-based (psychopathology) components, with correlations permitted between corresponding rater factors (e.g., youth survey and youth symptoms).

The final ESEM included 4 cognitive control factors (premeditation, activation/inhibitory control, and a task-based executive function factor) and 3 reward sensitivity factors (BAS, positive urgency, and sensation seeking). These factors were mediators between PGSs for internalizing, addiction, and neurodevelopmental traits and latent internalizing and externalizing psychopathology factors (Figure 1). Each PGS was tested in separate simple regressions and simultaneously in a multiple regression model to account for shared genetic variance.

All analyses used raw data in Mplus version 8.1 (59). Given that the sample includes siblings and we included all available data, we used the TYPE = COMPLEX option with family ID as a clustering variable, as this option has been found to correct for nonindependence due to siblings/twins (60). Weighted least squares mean and variance adjusted estimation was used for ordinal survey items, with an oblique Geomin rotation to allow factor correlations. Covariates included baseline age (highly collinear with year 2 age, r = 0.93), sex assigned at birth, and the first 10 genomic principal components. Model fit was evaluated using χ², root mean square error of approximation < 0.06, standardized root mean square residual < 0.08, and comparative fit index > 0.95 (61). Statistical significance was assessed through z tests with false discovery rate (FDR) correction (62) at α = .05, applied separately to 48 direct + indirect effects and 6 total effects (see the Supplemental Methods for details). Reported standardized regression coefficients can be squared to represent the proportion of unique variance explained in each outcome.

Results

PGSs Derived From Adult Psychopathology GWASs Predict Psychopathology in an Adolescent Sample

To examine whether PGSs derived from adult psychopathology GWAS discovery samples have predictive utility in adolescent samples, we first examined their total associations with the internalizing and externalizing psychopathology factors. As shown in Table 1, when considering each PGS in its own model (simple regressions), the total effects (c) of the internalizing, neurodevelopmental, and addiction PGSs to both internalizing psychopathology (c = 0.18, p_FDR < 10⁻⁴; c = 0.10, p_FDR < 10⁻⁴; and c = 0.05, p_FDR = .002, respectively) and externalizing psychopathology (c = 0.15, p_FDR < 10⁻⁴; c = 0.19, p_FDR < 10⁻⁴; and c = 0.08, p_FDR < 10⁻⁴, respectively) were significant.

Table 1.

Comparison of Total Indirect, Direct, and Total Effects in Simple and Multiple Regressions

Predictor	Internalizing Factor on PGS			Externalizing Factor on PGS
	Total Indirect Effects	Direct Effect(c′)	Total Effects(c)	Total Indirect Effects	Direct Effect(c′)	Total Effects(c)
Simple Regressions
Internalizing PGS	0.056∗	0.128∗	0.184∗	0.104∗	0.047∗	0.151∗
Addiction PGS	0.020∗	0.034	0.054∗	0.073∗	0.010	0.083∗
Neurodevelopmental PGS	0.063∗	0.038	0.101∗	0.148∗	0.043∗	0.191∗
Multiple Regression
Internalizing PGS	0.046∗	0.125∗	0.171∗	0.072∗	0.042∗	0.114∗
Addiction PGS	−0.011	0.000	−0.011	0.013	−0.012	0.001
Neurodevelopmental PGS	0.055∗	0.012	0.067∗	0.127∗	0.039∗	0.166∗

Sample size = 6667. Each row represents a separate analysis in the simple regressions. Estimates are standardized.

∗p_FDR < .05; 48 tests for indirect/direct effects: (3 PGSs × 7 ESEM outcomes × 2 psychopathology outcomes) + (3 PGSs × 2 psychopathology outcomes); 6 tests for total effects: 3 PGSs × 2 psychopathology outcomes.

ESEM, exploratory structural equation model; FDR, false discovery rate; PGS, polygenic score.

Reward Sensitivity and Cognitive Control Mediate Associations Between Genetic Liability for Psychopathology and Observed Psychopathology

Figure 2 and Table 2 show the specific indirect effects for each PGS to both the internalizing and externalizing factors. The associations between the internalizing PGS and the internalizing and externalizing factors were significantly mediated by latent variables related to both reward sensitivity and cognitive control.

Figure 2.

Path diagrams showing mediation results for each simple regression model. Ovals indicate latent variables, and squares indicate observed variables. Orange ovals indicate reward sensitivity latent variables, and green ovals indicate cognitive control latent variables. Red paths indicate significant mediation to externalizing factor, and purple paths indicate significant mediation to internalizing and externalizing factors. Strength of path is denoted by thickness, and statistical significance is denoted by solid lines. p_FDR < .05. Panel (A) shows the diagram for the simple regression model for the internalizing (INT) polygenic score (PGS). Panel (B) shows the diagram for the simple regression model for the neurodevelopmental (NEURO) PGS. Panel (C) shows the diagram for the simple regression model for the addiction (ADD) PGS. FDR, false discovery rate.

Table 2.

Comparison of Specific Indirect Effects Between Simple and Multiple Regression Models

Relationship	Mediator
	BAS(ab)	PU(ab)	SS(ab)	PRE(ab)	EF(ab)	ACT(ab)	INH(ab)
Simple Regressions
Predicting Internalizing
Internalizing PGS	0.000	0.000	0.005	0.000	0.001	0.015∗∗∗	0.035∗∗∗
Addiction PGS	0.000	0.000	−0.018∗∗	0.001	0.001	0.008∗∗	0.027∗∗∗
Neurodevelopmental PGS	0.001	0.000	−0.007	0.002	0.001	0.023∗∗∗	0.044∗∗∗
Predicting Externalizing
Internalizing PGS	0.000	0.001	0.001	0.001	0.006∗∗∗	0.036∗∗∗	0.060∗∗∗
Addiction PGS	0.002∗	0.001	−0.002∗	0.003	0.004∗∗	0.020∗∗	0.046∗∗∗
Neurodevelopmental PGS	0.003∗	0.001	−0.001	0.009∗∗∗	0.007∗∗∗	0.055∗∗∗	0.074∗∗∗
Multiple Regressions
Predicting Internalizing
Internalizing PGS	0.000	0.000	0.010	0.000	0.000	0.010∗∗∗	0.025∗∗∗
Addiction PGS	0.000	0.000	−0.019∗∗	0.000	0.000	−0.001	0.009
Neurodevelopmental PGS	0.000	0.000	−0.003	0.002	0.001	0.020∗∗∗	0.035∗∗∗
Predicting Externalizing
Internalizing PGS	−0.001	0.000	0.001	−0.001	0.004∗∗	0.025∗∗∗	0.042∗∗∗
Addiction PGS	0.002	0.001	−0.003	0.000	0.002	−0.003	0.015
Neurodevelopmental PGS	0.002	0.001	0.000	0.010∗∗∗	0.005∗∗	0.050∗∗∗	0.059∗∗∗

Sample size = 6667. Each column shows the standardized indirect effect of the psychopathology outcome on the PGS via the specified control or reward factor mediator.

∗p_FDR < .05, ∗∗p_FDR < .01, ∗∗∗p_FDR < .001; indirect and direct effects were FDR–corrected together for a total of 48 tests ([3 PGSs × 7 ESEM outcomes × 2 psychopathology outcomes] + [3 PGSs × 2 psychopathology outcomes]).

ab, a×b indirect effect; ACT, activation control factor; BAS, Behavioral Activation Scale factor; EF, executive function task factor; ESEM, exploratory structural equation model; FDR, false discovery rate; INH, inhibitory control factor; PGS, polygenic score; PRE, premeditation factor; PU, positive urgency factor; SS, sensation seeking factor.

The association between the internalizing PGS and observed psychopathology factors had the following multiple significant cognitive control mediators (ab): activation control (ab = 0.015, p_FDR < 10⁻⁴ to internalizing and ab = 0.036, p_FDR < 10⁻⁴ to externalizing); inhibitory control (ab = 0.035, p_FDR < 10⁻⁴ to internalizing and ab = 0.060, p_FDR < 10⁻⁴ to externalizing); and executive function (ab = 0.006, p_FDR = .001 to externalizing). No reward sensitivity factors significantly mediated the association between the internalizing PGS and psychopathology factors in the simple regression models.

The relationship between the neurodevelopmental PGS and internalizing was only significantly mediated by cognitive control factors: activation control (ab = 0.023, p_FDR < 10⁻⁴) and inhibitory control (ab = 0.044, p_FDR < 10⁻⁴). The relationship between the neurodevelopmental PGS and the externalizing factor was significantly mediated by several cognitive control factors (activation control ab = 0.055, p_FDR < 10⁻⁴; inhibitory control ab = 0.074, p_FDR < 10⁻⁴; executive function ab = 0.007, p_FDR < 10⁻⁴; and premeditation ab = 0.009, p_FDR < 10⁻⁴), as well as by a reward sensitivity factor (BAS ab = 0.003, p_FDR = .02).

The relationship between the addiction PGS and the internalizing factor was significantly mediated by cognitive control factors (activation control ab = 0.008, p_FDR = .002; and inhibitory control ab = 0.027, p_FDR < 10⁻⁴) as well as reward sensitivity (sensation seeking ab = −0.018, p_FDR = .001). The relationship between the addiction PGS and the externalizing factor was also significantly mediated by cognitive control factors (activation control ab = 0.020, p_FDR = .001; inhibitory control ab = 0.046, p_FDR < 10⁻⁴; and executive function ab = 0.004, p_FDR = .003) and reward sensitivity (sensation seeking ab = −0.002, p_FDR = .033; and BAS ab = 0.002, p_FDR = .035).

PGSs for Psychopathology Capture Some Genetic Signal Not Explained by Reward Sensitivity and Cognitive Control

The direct/total effect ratios (Figure 3, Table 1, and Figure S1) indicate the proportion of variance in the internalizing and externalizing factors unexplained by the mediators. For internalizing, the ratios were 0.70, 0.38, and 0.63 for the internalizing, addiction, and neurodevelopmental PGSs, respectively. Most of the variance explained by the internalizing and neurodevelopmental PGSs was direct, whereas most of the variance explained by the addiction PGS was mediated. In contrast, for externalizing, the ratios were 0.31, 0.23, and 0.12, indicating that most of the variance explained by all 3 PGSs was mediated.

Figure 3.

Indirect, direct, and total effects of neurodevelopmental, internalizing, and addiction risk-factor polygenic scores (PGSs) in the simple regression framework. Panel (A) shows a plot of effects for each PGS on psychopathology factors mediated by the cognitive control and reward sensitivity factors. Different mediators are shown in different colors and with different shapes. Panel (B) shows the total and direct effects of each PGS on the psychopathology factors. ADD, addiction risk-factor PGS; EXT, externalizing factor; INT, internalizing factor and internalizing PGS (PGSs are shown on the left of each label); NEURO, neurodevelopmental PGS.

The strongest direct effects (c′) in the simple regressions were from the internalizing PGS to both the internalizing and externalizing factors (c′ = 0.13, p_FDR < 10⁻⁴; c′ = 0.05, p_FDR < 10⁻⁴). The neurodevelopmental PGS showed a statistically significant direct effect on the externalizing factor (c′ = 0.04, p_FDR < 10⁻⁴) but not the internalizing factor (c′ = 0.04, p_FDR = .140). Neither of the direct effects of the addiction PGS on the psychopathology factors were statistically significant.

Comparison Between Simple Regression and Multiple Regression Models Suggests Pleiotropy

The correlations among the 3 PGSs (Table S11) ranged from r = 0.219 between the internalizing and neurodevelopmental PGSs to r = 0.330 between the neurodevelopmental and addiction PGSs. A comparison between the simple and multiple regression models provides information about how this overlap in genetic signal between the PGSs plays into shared behavioral or biological pathways across disorders.

The total effects of the internalizing, neurodevelopmental, and addiction PGSs on psychopathology in the multiple regressions largely mirrored the simple regressions. All PGSs remained significantly associated with internalizing and externalizing in the multiple regression models, except for the addiction PGS. While the addiction PGS showed significant total associations with both the internalizing and externalizing factors in the simple regressions, these effects were no longer significant in the multiple regression (internalizing: c = −0.011, p_FDR = .568; externalizing: c = 0.001, p_FDR = .981).

The indirect effects largely followed the same patterns of significance in the multiple regression as they did in the simple regression, with a few exceptions. First, while BAS significantly mediated the association between the neurodevelopmental PGS and externalizing in the simple regression, it was not significant in the multiple regression (ab = 0.002, p_FDR = .05). Second, the significant mediators of the association between the addiction PGS and the internalizing and externalizing factors in the simple regressions were largely attenuated in the multiple regression. The only indirect effect that remained significant in the multiple regression was the effect of sensation seeking (ab = −0.019, p_FDR = .001 on internalizing).

The direct/total effect ratios in the multiple regression followed the same trends as were seen in the simple regressions; indirect effects accounted for most of the genetic signal between the PGSs and observed psychopathologies for all pathways except for the internalizing PGS to the internalizing factor.

Discussion

Taken together, our results indicate that 1) PGSs derived from adult psychopathology factors predicted internalizing and externalizing factors in adolescents; 2) multiple cognitive control and reward sensitivity factors significantly mediated these associations; and 3) some relationships between PGSs and adolescent psychopathology were unrelated to these mediators. Notably, indirect effects accounted for most of the total effects across PGSs, except for the association between the internalizing PGS and the internalizing factor. These findings suggest that genetic risk manifests through several independent behavioral pathways, including cognitive control and reward sensitivity.

Predictive Utility of PGSs in an Adolescent Sample

Significant total effects from the PGSs to both the internalizing and externalizing psychopathology factors confirm the applicability of PGSs of adult psychopathology factors to an adolescent sample. The internalizing PGS explained ∼3% of variance in internalizing, while the neurodevelopmental and addiction PGSs explained 0.6% and 3% of variance in externalizing, respectively, estimates comparable with prior studies using PGSs to predict psychopathology (14). These modest effects likely reflect the limited variance captured by current GWASs, which are expected to increase as discovery sample sizes grow (3,14).

The total effects of the internalizing and neurodevelopmental PGSs on the internalizing and externalizing factors showed little attenuation going from simple regressions to multiple regressions. This pattern suggests that different PGSs (e.g., internalizing, neurodevelopmental) tap distinct portions of the genetic risk for internalizing and externalizing psychopathology in adolescence.

PGSs for Psychopathology Capture Genetic Signal Shared With and Unique From Reward Sensitivity and Cognitive Control

Mediation analyses offer insight into the mechanisms that these PGSs tap. Activation control and inhibitory control emerged as transdiagnostic risk factors, mediating relationships between multiple PGSs and both internalizing and externalizing psychopathology. Executive function and premeditation only significantly mediated pathways to externalizing. Notably, executive function mediated effects of internalizing and neurodevelopmental PGSs on externalizing psychopathology in the multiple regression (when all PGSs were included in the same model), suggesting that they capture unique genetic signals relevant to executive functioning.

Although executive function did not significantly mediate the internalizing PGS-internalizing pathway, the internalizing PGS was significantly associated with executive function. The literature on executive function and internalizing problems remains mixed; some studies have reported significant genetic associations (17,63), whereas others have found no genetic relationship or only phenotypic associations (17). These discrepancies may reflect differences in developmental stage—ranging from adolescence to adulthood—or variation in executive function measures (e.g., Trail Making Test, Backward Digit Span) not included here, which may differentially predict psychopathology.

Multiple cognitive control factors significantly mediated the neurodevelopmental PGS-externalizing association, highlighting that genetic influences on control processes contribute to externalizing symptomatology. This finding is consistent with previous evidence linking deficits in cognitive control and associated neural circuitry to both neurodevelopmental and externalizing disorders (64,65).

We observed attenuation of indirect and total effects for the addiction PGS when moving from a simple to a multiple regression framework. In the multiple regression model, no cognitive control factors remained significant. This is somewhat surprising given prior links between cognitive control and substance use at both phenotypic (66) and genetic levels (28). This may reflect shared genetic liability across PGSs, a type of pleiotropy that characterizes many psychiatric conditions (4,8). However, it is also important to consider that problematic alcohol use served as an indicator for both the addiction risk (14) and neurodevelopmental (29) factors, potentially increasing shared variance among associated PGSs.

Among the reward sensitivity factors, sensation seeking uniquely mediated the association between the addiction PGS and both internalizing and externalizing outcomes. The persistence of these effects in multiple regression models suggests that the addiction PGS captures genetic variance specifically linked to sensation seeking. Although the indirect effect from the internalizing PGS to sensation seeking was nonsignificant in the simple regression, it became marginally significant and negative in multiple regression, in contrast to the positive association observed for the addiction PGS. This negative relationship supports prior evidence that internalizing psychopathology is characterized by reduced reward sensitivity (67, 68, 69).

The internalizing PGS-internalizing pathway showed the highest proportion of direct/total effects, suggesting that reward sensitivity and cognitive control mediated less of this association than for externalizing. This pattern likely reflects stronger conceptual overlap between the internalizing GWAS phenotype and our internalizing factor, greater GWAS power enhancing predictive precision (70,71), and the omission of mediators capturing negative affect or related traits closely tied to internalizing risk (72,73). As our analyses focused on cognitive and reward-based pathways, future work incorporating affective or motivational mediators may better delineate mechanisms linking genetic risk to internalizing psychopathology.

Methodological Considerations

Self-report measures often predict psychopathology more strongly than tasks (74), although this pattern may partially reflect shared method variance with rater-based outcomes (75). Although we modeled rater factors to minimize such bias, some rater-based factors still showed stronger associations with psychopathology (22). The inclusion of PGSs in this study allowed us to examine this effect from another perspective. The PGSs, which are not rated by parents or youth, showed comparable associations with both task and rating measures of cognitive control, suggesting that ratings may not be better measures of psychopathology than tasks.

Limitations and Future Directions

First, it is important to consider interpretations of SEMs when considering our model of early behavioral risk factors mediating later psychopathology. As with all SEMs, our model tests how well the structure fits the underlying covariance matrix, but it is certainly true that other structures are possible.

Second, we modeled psychopathology using latent factors rather than relying solely on diagnostic categories. The KSADS, administered by trained interviewers, contributed approximately half of the indicators for the internalizing and externalizing latent variables. Although we also included survey-based measures (e.g., CBCL), we accounted for shared variance by incorporating rater factors at both the cognitive control/reward sensitivity and psychopathology levels.

Third, we did not include the psychotic and compulsive thought PGSs from the genomic SEM (29), as they did not improve prediction (see Tables S8 and S9). Researchers with targeted hypotheses may still benefit from examining these PGSs.

Fourth, the GWAS landscape is constantly evolving. Updated versions of the Grotzinger et al. (76) genomic SEM have removed problematic alcohol use as a neurodevelopment indicator and introduced a new substance use disorder factor without cross-loadings (77) of indicators from the substance use disorders factor to the neurodevelopmental factor (77). Because of the large overlap between included summary statistics in that model and the addiction risk-factor model (14,28), we chose to use the addiction risk-factor summary statistics instead (14) given that summary statistics from Paul et al. (29) and Hatoum et al. (14) allowed for integration with their studies, which both included phenome-wide association studies with these PGSs using the same ABCD sample.

Fifth, differences in LD structure and minor allele frequencies in different ancestral populations make translation of findings across ancestries difficult and not recommended (77). While multiple pipelines exist to increase the portability of PGSs across ancestral populations (78), a significant limitation of current methods, such as PRS-CSx (79, 80), still exists in the availability of ancestrally diverse GWASs (80). Furthermore, methods such as PRS-CSx have been shown to perform worse in the ABCD sample than European-derived PGSs applied to different ancestral groups (80). Thus, we ran PGSs in additional ancestral groups but maintain caution in interpreting the effects given that the summary statistics used in the discovery samples were based on GWASs of individuals with European-like ancestry.

Conclusions

Because PGSs capture genetic signal that is aggregated across the entire genome for an individual, multiple biological pathways may be captured in a single PGS. Our results suggest that PGSs for neurodevelopment, addiction, and internalizing psychopathology appear to capture multiple pathways relevant to cognitive control and reward sensitivity constructs. These findings elucidate the way that genetic risk factors for psychopathology may manifest in adolescence.