Recent advances in molecular genetics promise to elucidate mental health etiology in young people (Dick et al., Reference Dick, Barr, Cho, Cooke, Kuo, Lewis and Su2018; Thapar & Riglin, Reference Thapar and Riglin2020). Specifically, hundreds of individual risk alleles discovered in genome-wide association studies (GWASs) can be aggregated into polygenic risk scores (PRSs), which capture an individual' genetic vulnerability for a given disorder or trait. To date, PRSs have demonstrated an appreciable prediction of onset, severity, and developmental course of psychiatric conditions (Bogdan, Baranger, & Agrawal, Reference Bogdan, Baranger and Agrawal2018; Dick et al., Reference Dick, Barr, Cho, Cooke, Kuo, Lewis and Su2018; Thapar & Riglin, Reference Thapar and Riglin2020). Moreover, PRSs often show transdiagnostic associations with multiple mental health problems (Docherty et al., Reference Docherty, Moscati, Dick, Savage, Salvatore, Cooke and Riley2018; Krapohl et al., Reference Krapohl, Euesden, Zabaneh, Pingault, Rimfeld, Von Stumm and Plomin2016), suggesting that they might better capture vulnerability to broad, higher-order dimensions of psychopathology. However, the specificity of associations between a wide range of major psychopathology-related PRSs, and higher-order dimensions of childhood psychopathology, remain unexplored, complicating PRS applications in research and clinical practice.
Transdiagnostic and age-specific associations between PRSs and psychopathology
A large genetic overlap among psychiatric disorders is well documented in adults and children, and the overarching pattern of findings from GWASs indicates widespread pleiotropy, meaning that many variants influence more than one psychiatric disorder (Martin, Taylor, & Lichtenstein, Reference Martin, Taylor and Lichtenstein2017; Wray et al., Reference Wray, Lee, Mehta, Vinkhuyzen, Dudbridge and Middeldorp2014). Consequently, PRSs developed for one disorder or a trait have been found to significantly predict many other conditions (Docherty et al., Reference Docherty, Moscati, Dick, Savage, Salvatore, Cooke and Riley2018; Krapohl et al., Reference Krapohl, Euesden, Zabaneh, Pingault, Rimfeld, Von Stumm and Plomin2016). For example, a PRS for a personality trait neuroticism predicts depression, phobias, and multiple physical health outcomes in emerging adulthood (Docherty et al., Reference Docherty, Moscati, Dick, Savage, Salvatore, Cooke and Riley2018). Finally, due to changing manifestations of genetic risk across development, the associations between PRSs and youth psychopathology can show a different pattern than in adult samples (Belsky & Harden, Reference Belsky and Harden2019; Dick et al., Reference Dick, Barr, Cho, Cooke, Kuo, Lewis and Su2018; Thapar & Riglin, Reference Thapar and Riglin2020). For example, a recent study found that childhood-onset depression was associated with Depression-PRS, Schizophrenia-PRS, and ADHD-PRS, while adolescent-onset depression was only associated with Depression-PRS (Rice et al., Reference Rice, Riglin, Thapar, Heron, Anney, O'donovan and Thapar2018). Given the emerging evidence for age-specific associations between PRSs and psychopathology, studies focused on specific developmental periods are needed.
Hierarchical models of psychopathology can explicate PRSs
A major contributor to the precision of PRS summary statistics is the quality of psychiatric assessment in both the discovery GWAS and replication samples (Cai et al., Reference Cai, Revez, Adams, Andlauer, Breen, Byrne and Hamilton2020). Thus, the patterns of associations between PRSs and psychopathology outcomes can be affected by the known shortcomings of psychiatric measures, such as diagnostic unreliability, comorbidity among disorders, and heterogeneity within them (Waszczuk et al., Reference Waszczuk, Eaton, Krueger, Shackman, Waldman, Zald and Kotov2020). Limitations in psychiatric assessment can impact not only traditional case–control designs (Kotov et al., Reference Kotov, Krueger, Watson, Achenbach, Althoff, Bagby and Zimmerman2017), but also quantitative phenotypes based on diagnostic phenotypes (Newson, Hunter, & Thiagarajan, Reference Newson, Hunter and Thiagarajan2020).
The empirically derived models of psychopathology, such as the Hierarchical Taxonomy of Psychopathology, offer solutions to these problems by organizing mental health concerns into general and specific spectra (Caspi et al., Reference Caspi, Houts, Belsky, Goldman-Mellor, Harrington, Israel and Poulton2014; Forbes, Tackett, Markon, & Krueger, Reference Forbes, Tackett, Markon and Krueger2016; Kotov et al., Reference Kotov, Krueger, Watson, Achenbach, Althoff, Bagby and Zimmerman2017; Lahey et al., Reference Lahey, Applegate, Hakes, Zald, Hariri and Rathouz2012; Lahey, Krueger, Rathouz, Waldman, & Zald, Reference Lahey, Krueger, Rathouz, Waldman and Zald2016). In particular, the overall level of childhood psychopathology is encompassed by a general factor (‘p-factor’), which can be separated into more specific lower order factors, including the internalizing and externalizing dimensions. When applied within genetically informed framework, this hierarchical approach can clarify the specificity of existing PRSs in predicting childhood psychopathology.
Four previous studies have tested associations between PRSs and higher-order phenotypic psychopathology factors. First, Schizophrenia-PRS and Neuroticism-PRS were associated with the general factor of psychopathology in 16-year-old adolescents, with little evidence for associations with specific factors (Jones et al., Reference Jones, Heron, Hammerton, Stochl, Jones, Cannon and Linden2018). Second, Schizophrenia-PRS and ADHD-PRS were associated with the general factor in 7- and 13-year olds, with the ADHD-PRS and Depression-PRS also showing independent associations with specific factors (Riglin et al., Reference Riglin, Thapar, Leppert, Martin, Richards, Anney and Lahey2019). A third study in 9–12-year-old children found that ADHD-PRS was associated both with the general factor and the hyperactivity/impulsivity symptoms (Brikell et al., Reference Brikell, Larsson, Lu, Pettersson, Chen, Kuja-Halkola and Martin2018). Finally, a recent study derived a polygenic general psychopathology factor based on eight psychiatric PRSs, and found that it predicted the p-factor across development (Allegrini et al., Reference Allegrini, Cheesman, Rimfeld, Selzam, Pingault, Eley and Plomin2020). Overall, these analyses demonstrated that in developmental samples, selected psychiatric PRSs appear to capture largely – but not exclusively – broad genetic influences at the level of the general factor of psychopathology. However, these studies had a limited scope, as tested only a small set of PRSs, thus not considering the genetic risk scores available for many psychiatric conditions and relevant traits.
Current study
Our group has recently delineated and validated a comprehensive higher-order structure of childhood psychopathology in a large sample of 9–10-year-old children from the Adolescent Brain Cognitive Development (ABCD) Study (Michelini et al., Reference Michelini, Barch, Tian, Watson, Klein and Kotov2019). A simplified summary of the derived structure is presented in Fig. 1a, with the highest level in the hierarchy consisting of the general factor, and at the lowest level consisting of five moderately correlated factors: externalizing, internalizing, neurodevelopmental, somatoform, and detachment. The aim of the current study was to build directly on this phenotypic work in the ABCD sample, to explore the associations between a wide-range of psychiatric PRSs and the higher-order dimensions of childhood psychopathology. Specifically, we tested which level of the psychopathology hierarchy best captures contributions of each PRS.
Fig. 1. (a) Graphical representation of the 1- and 5-factor structure of the CBCL scales in the ABCD Sample; (b) Proportion of total associations between PRSs and CBCL psychopathology due to the factors.
Notes: (a) is a graphical representation of the CBCL structure from Michelini et al. (Reference Michelini, Barch, Tian, Watson, Klein and Kotov2019). The intermediate 2-, 3-, and 4-factor solutions reported in the original paper are not the focus of the current study and are omitted from the figure. The factor loadings come from a confirmatory factor analysis in the analytic subsample used in the current study. The model fit was appropriate in the current subsample: root mean square error of approximation (RMSEA) = 0.031 (95% CI 0.030–0.031), standardized root mean square residual (SRMR) = 0.09, Comparative Fit Index (CFI) = 0.90, Tucker−Lewis Index (TLI) = 0.90. In (b), only the 15 PRSs that showed significant associations with CBCL psychopathology factors in regression models are depicted. The total R 2 reflects associations between PRSs and CBCL 1- and 5- factors, exclusive of 10 PCs.
Methods
Participants
Participants came from the ABCD study (Garavan et al., Reference Garavan, Bartsch, Conway, Decastro, Goldstein, Heeringa and Zahs2018). The total sample consists of over 11 000 children aged 9–10 years old recruited from 21 locations across the US. Inclusion criteria were the child' age and attendance of selected schools within the catchment areas, with the exception for twin recruitment, which relied on birth registers. Exclusion criteria were the lack English proficiency; major sensory, medical, neurological or psychiatric conditions; inability to complete an MRI scan; and gestational age <28 weeks or birthweight <1200 g. The total sample was 51.0% White, 21.4% Hispanic, 15.2% African American, 2.3% Asian, and 10.0% multiracial or from other backgrounds. The household annual income was 30.5% under $50 000, 28.1% between $ 50 000 and $ 100 000; and 41.3% over $ 100 000. The sample consisted of 73.3% families with parents either married or living together, and 58.9% with at least one parent holding a university-level degree. Overall, the sample approached socio-demographic characteristics of the US population, despite not being nationally representative (Compton, Dowling, & Garavan, Reference Compton, Dowling and Garavan2019). All procedures were approved by a central Institutional Review Board at the University of California, San Diego. Parents provided written informed consent, and children provided assent prior to study participation.
The current study uses data from the Baseline ABCD 2.0 data release (NDAR-https://doi.org/10.15154/1503209). The analytic sample consisted of unrelated children, thus in case of genetic or self-reported familial relatedness, only one child was randomly selected into the analyses (N = 1422 excluded). Moreover, only participants of European Ancestry were included in analyses, with ancestry evaluated using genetic data (N = 4458 excluded). This is because PRSs are sensitive to ancestry (Martin et al., Reference Martin, Kanai, Kamatani, Okada, Neale and Daly2019), thus analyses were limited to individuals with ancestry matching the discovery GWASs. Overall, after the genotyping quality control (N = 62 excluded), the final analytic sample for which phenotype and genotype data was available totaled N = 4717. The mean age of the analytic sample was 9.92 years old (s.d. = 0.62, range = 9.00–10.92 years old) and 47.1% of participants were female.
The Child Behavior Checklist
The Child Behavior Checklist (CBCL) (Achenbach & Rescorla, Reference Achenbach and Rescorla2001) was used to assess childhood psychopathology. Parents rated child' emotional and behavioral problems occurring in the past 6 months on a 119-item, 3-point scale questionnaire during an in-person visit. Items have been aggregated into higher-order dimensions that have been previously empirically delineated and validated in the ABCD sample, using an exploratory principal component analysis with an oblique (geomin) rotation (Michelini et al., Reference Michelini, Barch, Tian, Watson, Klein and Kotov2019), see Fig. 1a. Specifically, the current analyses focused only on the 1 and 5-factor solutions, and examined whether these factors demonstrate a good fit in the analytic subsample using a confirmatory factor analysis.
Genotyping
Genotyping of saliva samples stored at the Rutgers University Cell and DNA Repository was performed using the Smokescreen Array (Baurley, Edlund, Pardamean, Conti, & Bergen, Reference Baurley, Edlund, Pardamean, Conti and Bergen2016), according to protocols of the manufacturer. Genotypes were imputed on the Michigan Imputation Server pipeline v1.2.4, using the Haplotype Reference Consortium reference panel (McCarthy et al., Reference McCarthy, Das, Kretzschmar, Delaneau, Wood, Teumer and Sharp2016). Before imputation, the genotypes were filtered for ambiguous strand orientation, missingness rate>5% (by marker exclusion, then by individual), Hardy−Weinberg equilibrium violation (p < 10−6), sex mismatch (‘sex check’ function for X chromosome homozygosity estimate), and non-European ancestry (principal component analysis against the reference panel from the 1000 Genomes data). After imputation, the SNPs were excluded for imputation R 2 < 0.5, average call rate below 90% and minor allele frequency below 0.1%. PLINK was used to handle genetic data and perform quality control (Purcell et al., Reference Purcell, Neale, Todd-Brown, Thomas, Ferreira, Bender and Daly2007). The pi_hat > 37.5% was used for the identity by descent exclusion (one sample per pair was removed at random), and samples of less than 80% genetic European ancestry were excluded. Genotype imputation was performed on 487 562 SNPs, resulting in 19 519 349 SNPs after quality control which were used for the final polygenic risk scoring.
Polygenic risk scores
Polygenic risk scores were created for 22 phenotypes relevant to psychopathology, using summary statistics from GWAS discovery samples listed in Table 1. To minimize PRS standard error, only phenotypes from GWASs that included over 100 000 participants were selected. As an exception, we included PRSs for ADHD (total N = 55 374) and autism spectrum disorder (total N = 46 350), due to their unique relevance to developmental populations and neurodevelopmental psychopathology. PRSs were computed, using the PRSice 2.0 software (Euesden, Lewis, & O'Reilly, Reference Euesden, Lewis and O'Reilly2015), with r 2 = 0.1 threshold of clumping, by aggregating genetic variants up to varying thresholds of significance, weighted by the associations in the GWAS sample. Our primary analysis used a full list of SNPs and weights a priori, to minimize multiple testing (p-value threshold = 1). See online Supplementary Table S1 for bivariate associations between PRSs, and online Supplementary Table S2 for sensitivity analyses at other thresholds: 0.01, 0.05, 0.10, and 0.50. The first 10 genetic ancestry principle components (PCs) were obtained as measures of population stratification.
Table 1. Discovery GWASs for the PRSs tested in the current study
Statistical approach
All variables were standardized prior to analyses. A false detection rate (FDR) of 5% was used to adjust for multiple comparisons. First, bivariate associations of PRSs with CBCL factors were tested using partial correlations, controlling for the first 10 PCs of the population structure. Second, bivariate partial correlations of PRSs with dimensions from the 5-factor solution were repeated, this time additionally controlling for the general factor variance, to capture specific associations. Post hoc power analyses were computed using the AVENGEME script (Dudbridge, Reference Dudbridge2013).
Third, in order to test which level of the phenotypic hierarchy (1 v. 5-factors) is most strongly associated with each PRS, the R 2 change test in hierarchical linear regression models was used, adjusted for the number of degrees of freedom. Specifically, in each test, a PRS constituted a dependent variable. In the first block, 10 ancestry PCs were included to estimate baseline R 2. In the second block, the general factor from the 1-factor solution was added, to estimate the incremental change in R 2, and thus determine whether the general factor is significantly associated with each PRS. In the third block, all factors from the 5-factor solution were added simultaneously, with the change in R 2 informing whether this level of the phenotypic hierarchy provides incremental association with PRS, over and above the PCs and the general factor. Finally, for each PRS, the proportion of total R 2 due to associations with 1 v. 5-factors was calculated.
Results
Bivariate partial correlations with the general factor
The bivariate partial correlations indicated that 13 out of 22 tested PRSs were significantly associated with the general factor of psychopathology (Table 2). The largest associations were observed for Chronic Multisite Pain-PRS (r = 0.098, s.d. = 0.014, p < 0.001), ADHD-PRS (r = 0.079, s.d. = 0.015, p < 0.001), Depression-PRS (r = 0.078, s.d. = 0.015, p < 0.001), Smoking-PRS (r = −0.077, s.d. = 0.015, p < 0.001), Neuroticism-PRS (r = 0.075, s.d. = 0.014, p < 0.001), and Educational Attainment-PRS (r = −0.064, s.d. = 0.014, p < 0.001). The other PRSs significantly associated with the general factor were, in the order of effect size, Chronic back pain-PRS, Insomnia-PRS, Number of Sexual Partners-PRS, Autism-PRS, BMI-PRS, PTSD-PRS, and Disinhibition-PRS (r = 0.039-0.056). The pattern of associations was generally consistent across other p-value thresholds, although the most restrictive threshold (p = 0.01) yielded the weakest associations (online Supplementary Table S2).
Table 2. Bivariate partial correlations between PRSs and CBCL factors
Notes: All significant results at 5% FDR multiple testing correction are in bold. All correlations adjust for the first 10 ancestry PCs.
Bivariate partial correlations with the five specific factors
The associations between PRSs and the five specific factors emerged most frequently for the externalizing, neurodevelopmental, and somatoform factors, in part due to the largest number of PRSs representing the externalizing domain (Table 2). However, after adjusting for the general factor, none of the associations with the somatoform factor remained significant, and only ADHD-PRS was significantly associated with the neurodevelopmental factor (r = 0.062, s.d. = 0.026, p < 0.001) (Table 3). Conversely, the externalizing factor remained independently associated with Smoking-PRS, Adventurousness-PRS, Disinhibition-PRS, Educational Attainment-PRS, and Number of Sexual Partners-PRS (|r| = 0.040–0.058).
Table 3. Bivariate partial correlations between PRSs and five lower-order CBCL factors, adjusted for the general factor
Notes: All significant results at 5% FDR multiple testing correction are in bold. All correlations adjust for the first 10 ancestry PCs and the general factor.
The internalizing and detachment factors demonstrated notably fewer associations with PRSs (Table 2). However, after adjusting for the general factor, many PRSs showed protective effects on the internalizing dimension (Table 3). For example, the internalizing factor was negatively associated with Number of Sexual Partners-PRS (r = −0.051, s.d. = 0.020, p = 0.001), and positively associated with Educational Attainment-PRS and Intelligence-PRS (r = 0.056–0.068). Similarly, after adjusting for the general factor, the detachment dimension was associated with lower genetic vulnerability to Alcohol, Disinhibition, and Adventurousness (|r| = 0.045 to 0.040, Table 3).
Finally, six PRSs showed no associations with any specific factor, beyond their significant associations with the general factor: Autism-PRS, Chronic Back Pain-PRS, Insomnia-PRS, Depression-PRS, Neuroticism-PRS, and PTSD-PRS (Table 3). Conversely, while not significantly associated with the general factor, five PRSs showed only associations with specific factors: Adventurousness-PRS with externalizing, internalizing, and detachment factors; Intelligence-PRS with internalizing factor; Knee Pain-PRS with internalizing factor; and two Alcohol-PRSs with detachment factor.
Mapping the hierarchical structure to PRSs
The regression analyses found that five PRSs – Chronic Back Pain-PRS, Insomnia-PRS, Depression-PRS, Neuroticism-PRS, and PTSD-PRS − were associated only with the general factor, so that the addition of the dimensions from the 5-factor solution did not significantly increase the association between PRS and psychopathology (Table 4). Between 66.7 and 85.7% of total associations between these five PRSs, as well as Chronic Multisite Pain-PRS, and childhood psychopathology was captured by the general factor (Fig. 1b). Autism-PRS, ADHD-PRS, Number of Sexual Partners-PRS, and Smoking-PRS contributed significantly and approximately equally to the general and specific forms of psychopathology.
Table 4. Summary of incremental associations between PRS and CBCL general factor and five lower-order factors
Notes: Significant ΔR 2 between regression blocks are demarked with an asterisk, after 5% FDR multiple testing correction.
For full hierarchical regression results, see online Supplementary Table S2.
Conversely, BMI-PRS, Disinhibition-PRS, Educational Attainment-PRS, Adventurousness-PRS, and Intelligence-PRS contributed more to the five dimensions than to the general factor, with 0−40.0% of total genetic associations captured by the general factor (Table 4 and Fig. 1b). In fact, the Adventurousness-PRS and Intelligence-PRS were not associated with the general factor, but there was a significant increase in the association with psychopathology once the five dimensions were added.
Discussion
The current study delineated a pattern of associations between a wide range of PRSs, and empirically derived, hierarchical factors of childhood psychopathology in the ABCD study. Specifically, the PRSs for internalizing and somatoform problems predominantly captured non-specific genetic vulnerability to psychopathology in childhood. Conversely, PRSs for a range of externalizing problems, ADHD, BMI, educational attainment, and intelligence contributed more to specific than to general psychopathology, most notably to the externalizing factor. Overall, the findings indicate that some genetic effects captured by major PRSs are non-specific, but the balance of general v. specific effects varied substantially across PRSs.
We found that a wide range of PRSs were associated with the general factor of psychopathology. This is in line with twin and molecular genetic studies showing widespread pleiotropy underpinning psychiatric conditions in youth and adults (Martin et al., Reference Martin, Taylor and Lichtenstein2017; Wray et al., Reference Wray, Lee, Mehta, Vinkhuyzen, Dudbridge and Middeldorp2014). In particular, PRSs derived for internalizing and somatoform phenotypes were among the most strongly associated with the general factor. There are three possible explanations for this. First, the CBCL-based general factor in childhood might be predominantly characterized by internalizing and somatoform variance (e.g. distress, fears, pain). However, in the current higher-order model, the general factor did not have a disproportionate contribution from the internalizing or somatoform spectrum, as evidenced by Fig. 1a loadings and the original model in Michelini et al. (Reference Michelini, Barch, Tian, Watson, Klein and Kotov2019). Second possible explanation is that GWASs for internalizing and somatoform conditions were particularly saturated with general psychopathology variance, resulting in PRSs that captured little specificity. This is consistent with prior findings that genetic vulnerability to depression was the strongest marker of the general genetic factor (Selzam, Coleman, Caspi, Moffitt, & Plomin, Reference Selzam, Coleman, Caspi, Moffitt and Plomin2018; Waldman, Poore, Luningham, & Yang, Reference Waldman, Poore, Luningham and Yang2020). Third, genetic vulnerability for internalizing and somatoform difficulties in adulthood manifests primarily as a general risk for psychopathology in childhood.
The results demonstrated incremental validity of investigating associations between PRSs and specific factors of psychopathology. We found that Adventurousness-PRS, Disinhibition-PRS, Number of Sexual Partners-PRS, Smoking-PRS, and Education Attainment-PRS captured genetic vulnerability specific to the externalizing dimension in youth, over and above the genetic vulnerability to general psychopathology. Moreover, the ADHD-PRS predicted the neurodevelopmental factor, independently of its significant association with the general factor, in line with the pattern of results reported by Riglin et al. (Reference Riglin, Thapar, Leppert, Martin, Richards, Anney and Lahey2019) and Brikell et al. (Reference Brikell, Larsson, Lu, Pettersson, Chen, Kuja-Halkola and Martin2018). In fact, more than half of the total associations between these PRSs, and childhood psychopathology, were due to the specific factors. This suggests that GWASs for externalizing problems and ADHD might be better able to capture the narrow phenotypes of interest, resulting in more precise PRSs, perhaps because target phenotypes are characterized by observable behavioral symptoms. This interpretation is also consistent with prior finding that molecular genetic markers of antisocial behavior and substance use were only moderately associated with the general genetic factor (Waldman et al., Reference Waldman, Poore, Luningham and Yang2020). Notably, the ADHD-PRS emerged as the most predictive PRS in the current study, accounting for a total of 1.4% of variance in CBCL factors. This suggests that ADHD-PRS might be one of the strongest currently available PRSs for capturing genetic vulnerability to psychopathology in youth.
Accounting for the genetic vulnerability to general psychopathology allowed to uncover potential protective genetic effects. For example, Adventurousness-PRS and Disinhibition-PRS were negatively associated with the detachment dimension, which is consistent with withdrawn behaviors captured by this factor. However, negative associations observed for the internalizing spectrum after accounting for the general factor, for example with the Educational Attainment-PRS, are more difficult to interpret. The reverse pattern of associations between the residual internalizing factor and a wide range of vulnerability factors has been reported previously, and might constitute a statistical artifact, such as too little variance remaining in the specific factor (Bornovalova, Choate, Fatimah, Petersen, & Wiernik, Reference Bornovalova, Choate, Fatimah, Petersen and Wiernik2020; Watts, Poore, & Waldman, Reference Watts, Poore and Waldman2019). Overall, the potential protective effects of internalizing PRSs should be interpreted with caution. Finally, none of the PRSs tested in this study captured the variance unique to the somatoform factor, independently of the general factor. This may indicate that genetic risk for somatoform phenotypes is mediated by the general factor in childhood.
Overall, the pattern of results is in line with the evidence that, while some polygenetic vulnerability is broad and transdiagnostic, a significant proportion of polygenetic risk is specific to narrower psychiatric constructs, underscoring the added benefit of fine-grained modeling of psychopathology (Waszczuk et al., Reference Waszczuk, Eaton, Krueger, Shackman, Waldman, Zald and Kotov2020). The hierarchical genetic architecture has previously been demonstrated in family and twin studies (Lahey et al., Reference Lahey, Krueger, Rathouz, Waldman and Zald2016; Pettersson, Larsson, & Lichtenstein, Reference Pettersson, Larsson and Lichtenstein2016; Waldman, Poore, van Hulle, Rathouz, & Lahey, Reference Waldman, Poore, van Hulle, Rathouz and Lahey2016), and more recently using molecular genomic structural equation modeling (Grotzinger et al., Reference Grotzinger, Rhemtulla, de Vlaming, Ritchie, Mallard, Hill and Deary2019; Waldman et al., Reference Waldman, Poore, Luningham and Yang2020). The results have several implications for the translation of PRS for research use.
First, the comprehensive pattern of associations between PRSs and hierarchical psychopathology was obtained in order to spotlight the varied degrees of predictive specificity of the currently available PRSs. The degree to which genetic risk is general or specific is relevant to the conduct and interpretation of applied genetic research, as it can confound mechanistic interpretations. For example, current internalizing and somatoform PRSs (i.e. Depression-PRS, Neuroticism-PRS, Insomnia-PRS, PTSD-PRS, Chronic Multisite Pain-PRS, Chronic Back Pain-PRS) need to be applied with an understanding that they largely capture risk factors for broad psychopathology rather than genetic mechanisms specific to a given disorder.
Second, the results also have implications for the design of future GWAS. As differential genetic discovery is expected to emerge at each level of the phenotypic structure, future GWASs on phenotypes at different levels of the hierarchy could help create more precise PRSs. Recently, Cai et al. (Reference Cai, Revez, Adams, Andlauer, Breen, Byrne and Hamilton2020) empirically demonstrated that imprecise, minimal phenotyping yields Depression-PRS that broadly predicts psychopathology, while narrow phenotypic definitions markedly increase the prediction specificity of Depression-PRS. Thus, future discovery GWASs should explicitly calibrate target phenotypes to develop transdiagnostic and p-value PRSs alongside disorder-specific PRSs.
Limitations
The current study has several strengths, including a comprehensive assessment of child psychopathology in a large, well-characterized, and genotyped cohort. Nonetheless, some limitations are notable. First, analyses were restricted to participants of European Ancestry, who constitute only about half of the ABCD cohort, considerably curbing the representativeness of the current findings. The PRSs have been developed in discovery GWAS samples of European ancestry and are known to perform poorly in individuals from other ancestries (Martin et al., Reference Martin, Kanai, Kamatani, Okada, Neale and Daly2019). One improvement would be additional discovery GWASs including ancestry admixture. Second, the PRS−phenotype association is influenced by similarities between the discovery GWAS and the test sample, including demographic characteristics and instruments used to collect data in both samples. Thus, the pattern of phenotypic manifestation of PRSs reported in the current study might not replicate in other adolescent samples, and when GWASs based on larger and/or developmental discovery samples become available. The sample characteristics, specifically the lack of severe psychopathology in the ABCD cohort, might partially explain the unexpected null findings for Bipolar Disorder-PRS and Schizophrenia-PRS. Furthermore, we selected PRSs to ensure adequate power and nearly all of them predicted psychopathology dimensions (Dudbridge, Reference Dudbridge2013). However, strength of discovery samples differed substantially among these PRS, thus the precision in evaluating specificity of some PRSs was stronger than others. In particular, ADHD-PRS, Autism-PRS, Disinhibition-PRS, Number of Sexual Partners-PRS, and PTSD-PRS were estimated to have power somewhat lower than 80%, but nonetheless demonstrated their utility for present analyses by predicting psychopathology in ABCD data.
Third, the phenotypic structure was obtained using cross-sectional data and its stability or measurement invariance over time has not yet been established. Fourth, as typical for this age group, youth psychopathology was assessed only via parent-report. There are different strengths and limitations associated with self- and parent-reports (Bird, Gould, & Staghezza, Reference Bird, Gould and Staghezza1992; Jensen et al., Reference Jensen, Rubio-Stipec, Canino, Bird, Dulcan, Schwab-Stone and Lahey1999), with informants showing only moderate agreement due to their different perspectives (Goodman, Reference Goodman2001; Stanger & Lewis, Reference Stanger and Lewis1993). Nonetheless, twin studies suggest that different informants appear to measure a largely common genetic liability (Allegrini et al., Reference Allegrini, Cheesman, Rimfeld, Selzam, Pingault, Eley and Plomin2020; Bartels, Boomsma, Hudziak, van Beijsterveldt, & van den Oord, Reference Bartels, Boomsma, Hudziak, van Beijsterveldt and van den Oord2007; Merwood et al., Reference Merwood, Greven, Price, Rijsdijk, Kuntsi, McLoughlin and Asherson2013). Finally, PRSs capture only a small proportion of genetic influences on psychiatric disorders, and account for too little variance to be clinically informative. We were not able to account for the associations with the remaining genetic liability that is not captured by the PRSs.
Conclusions
The current study investigated phenotypic manifestations of polygenic risk in a developmental sample. Genetic vulnerability captured by the internalizing and somatoform PRSs was largely associated with the general factor of psychopathology in childhood, indicating that these PRSs predominantly capture transdiagnostic genetic effects. Moreover, some externalizing PRSs showed independent and incremental associations with the specific psychopathology dimensions, most notably with the externalizing factor. Likewise, ADHD-PRS was uniquely associated with the neurodevelopmental factor, over and above the general factor. Overall, specific genetic effects were often at least as large as general effects, but varied substantially across PRSs. The results demonstrate that higher-order models of psychopathology can help to explain patterns of PRS prediction in childhood samples.
Acknowledgements
This work was supported by the National Institute of Mental Health (A.D., grant number K01MH093731) and the Brain & Behavior Research Foundation (A.D. & A.S.). Data used in the preparation of this article were obtained from the Adolescent Brain Cognitive Development (ABCD) Study (https://abcdstudy.org), held in the NIMH Data Archive (NDA). This is a multisite, longitudinal study designed to recruit more than 10 000 children age 9–10 and follow them over 10 years into early adulthood. The ABCD Study is supported by the National Institutes of Health and additional federal partners under award numbers U01DA041022, U01DA041028, U01DA041048, U01DA041089, U01DA041106, U01DA041117, U01DA041120, U01DA041134, U01DA041148, U01DA041156, U01DA041174, U24DA041123, U24DA041147, U01DA041093, and U01DA041025. A full list of supporters is available at https://abcdstudy.org/federal-partners.html. A listing of participating sites and a complete listing of the study investigators can be found at https://abcdstudy.org/scientists/workgroups/. ABCD consortium investigators designed and implemented the study and/or provided data but did not necessarily participate in analysis or writing of this report. This manuscript reflects the views of the authors and may not reflect the opinions or views of the NIH or ABCD consortium investigators. The ABCD data repository grows and changes over time. The ABCD data used in this report came from NDAR-https://doi.org/10.15154/1503209.
Conflict of interest
The authors declare no conflicts of interest.
Supplementary material
The supplementary material for this article can be found at https://doi.org/10.1017/S0033291721003639
