The copy number variant architecture of child psychopathology and cognitive development in the ABCD^® study

Abstract

Objective:

Late childhood is a crucial developmental period for domains of mental health and cognition affected in individuals with psychiatric disorders. While common single nucleotide polymorphisms explain a large proportion of inherited genetic risk, structural variations including copy number variants (CNVs) play a significant role in the genetic architecture of neurodevelopmental disorders. Despite their importance, the relevance of CNVs to child psychopathology and cognitive function in the general population remains underexplored.

Methods:

Using two algorithms for CNV detection, we identified duplications and deletions across a cohort of 11,876 individuals from the ABCD study. Quality control procedures considered array log R ratio and B allele frequency profiles, CNV size, agreement between the two algorithms, and specific genomic location of CNVs. CNVs meeting quality control were used to identify regions associated with quantitative measures of broad psychiatric symptom domains and cognitive functioning. Additionally, CNV risk scores, reflecting the aggregated burden of genetic intolerance to inactivation and dosage sensitivity, were calculated to assess cumulative impact on overall and dimensional psychiatric and cognitive phenotypes.

Results:

Across 8,564 individuals passing quality control, 4,111 individuals carried 5,760 autosomal CNVs. Although no CNV regions reached significance after applying strict multiple testing correction, 16 CNV regions showed associations with psychopathology and cognitive development at an uncorrected genome-wide significance level. A duplication at 14q11.2 showed the strongest association with attentional psychopathology. Moreover, individuals carrying CNVs previously associated with neurodevelopmental disorders exhibited greater impairment in social functioning and cognitive performance across fluid intelligence, working memory and processing speed. Notably, higher CNV risk scores were significantly correlated with greater attention problems and cognitive impairment across multiple domains (fluid intelligence, attention, working memory, flexible thinking, and processing speed).

Conclusions:

Our findings shed light on the contributions of CNVs to interindividual variability in complex traits related to neurocognitive development and child psychopathology.

Introduction

Neurodevelopmental disorders are characterized by a spectrum of mental health issues, developmental delays, and deficits in diverse cognitive functions(1, 2). Genetic factors are known to play a crucial role in these disorders(3, 4) such as autism spectrum disorder(5, 6), attention-deficit/hyperactivity disorder(7, 8) and psychosis spectrum disorders(9). Population-based genetic studies have revealed robust associations between common variants, such as single-nucleotide polymorphisms (SNPs), and the pathogenesis of these conditions(10, 11). The advent of genome-wide association scans has shed light on the polygenic nature of neurodevelopmental disorders. However, the contribution of SNPs with subtle genetic effects does not fully account for the variance in these complex phenotypes due to genetic factors (12). In addition to common genetic variants, other types of genetic variations including copy number variants (CNVs) – a form of genomic structural rearrangement involving duplications and deletions spanning at least 1000 base pairs throughout the human genome(13) – are increasingly recognized as crucial components in elucidating the detailed genetic architecture underlying the etiology of neurodevelopmental disorders(14).

While the large majority of CNVs are benign, clinical and genetic epidemiological investigations have consistently demonstrated that individuals harboring pathogenic CNVs often exhibit early-onset symptoms of mental illness and are more predisposed to transitioning into clinically diagnosable psychiatric disorders(15). For example, approximately one quarter of individuals with the 22q11.2 deletion develop schizophrenia(16, 17). Large-scale case-control studies have identified a spectrum of genomic duplications and deletions associated with increased genetic susceptibility to neurodevelopmental disorders, including schizophrenia, attention-deficit/hyperactivity disorder, and autism spectrum disorder(5, 6, 18). However, these investigations have predominantly focused on identifying genetic correlates within the confines of a categorical diagnostic framework, while it is increasingly acknowledged that psychopathological manifestations and associated cognitive deficits exist along a continuum beyond traditional diagnostic boundaries(19–21). To date, there remains a dearth of research exploring the CNV architecture underlying dimensions of psychopathology and cognitive functions during childhood.

The concept of the CNV risk score has emerged to estimate the tolerance to loss-of-function mutations and dosage sensitivity of genes by deletions or duplications(22). CNV risk scores aim to measure the cumulative risk from deletions or duplications of large genomic segments. Different types of genetic variants exhibit cumulative effects, contributing to overall phenotypic variability(12). Recently, we observed that higher CNV risk scores were significantly correlated with poor cognitive performance and greater psychopathology in youth(23). These findings underscore the potential of CNV risk scores to detect genetic associations with neurodevelopmental outcomes in childhood and adolescence.

In the present study, we conducted a comprehensive exploration of the CNV architecture underlying dimensions of psychopathology and cognitive phenotypes within the Adolescent Brain Cognitive Development (ABCD^®) Study. This prospective study comprises a community sample of N=11,876 children aged 9–10 years old, who were not ascertained for mental health problems(24). Our investigation proceeded in several stages. First, CNVs for each individual were identified using two algorithms, PennCNV(25) and QuantiSNP(26), leveraging genome-wide SNP array data. Second, CNVs that overlapped between the two algorithms were utilized in a genome-wide CNV association scan, aiming to identify specific CNV regions (CNVRs) associated with psychopathology and cognition independently. Third, we examined whether individuals carrying CNVs associated with neurodevelopmental disorders(22) exhibited poorer cognitive performance and greater psychopathology. Finally, we explored the association of CNV risk score with subdomains of psychopathology and cognition. Cumulatively, the present study sheds significant light on the contribution of CNVs to interindividual variability in complex traits critical to neurodevelopmental outcomes.

Methods

Sample and CNV quality control

The ABCD^® study acquires genetic data by the Chemagen bead-based/Chemagic STAR DNA Saliva4k Kit (CMG-1755-A) at the Rutgers University Cell and DNA Repository(27). Genotypes were ascertained using the Affymetrix National Institute on Drug Abuse SmokeScreen Array(28), and imputed using SHAPEIT3(29) and impute2(30). Data Analysis, Informatics, and Resource Center in ABCD^® performed these initial quality control steps in alignment with established best practices(31), including recommendations from the Ricopili pipeline(32). Specifically, the primary genotype quality control steps included excluding samples with a genotype call missing rate >20% and variants with a missing rate >10%. Samples with any implausible numbers of third-degree relatives were also removed to maintain data integrity. ABCD^® compiled log R ratio and B allele frequency profiles for each individual based on quality-controlled genotype data. Following these quality control processes, 11,088 individuals were retained for downstream analyses. These genotype data were subsequently used to call CNVs without additional filtering, as the curated genotype data in ABCD^® study meets rigorous quality control standards.

To call CNVs, we downloaded the compiled log R ratio and B allele frequency profiles for each SNP marker from the ABCD^® portal (Data Release 3.0). Multiple sample log R ratio and B allele frequency matrix files, paired with sample and probe information files were split into single sample log R ratio and B allele frequency files. Population frequency of B allele and GC base content model definition files specific to the Affymetrix SmokeScreen Array were compiled. CNV calling was performed following established procedures detailed in our previous publications (Figure 1)(23, 33). Specifically, we employed hidden Markov models to call CNVs using both PennCNV(25) and QuantiSNP(26) algorithms. PennCNV utilizes multiple sources of information, including the log R ratio and B allele frequency at each SNP marker, the distance between neighboring SNPs, and adjusts for “genomic waves” using a regression model based on GC content(25). On the other hand, QuantiSNP utilizes an objective Bayes approach, leveraging the log R ratio and B allele frequency for each SNP marker(26). Both algorithms have demonstrated sensitivity in detecting deletions and duplications in the human genome across genetic studies spanning the past two decades(9, 34, 35).

Figure 1. Schematic of the pipeline for quality control of CNVs in the ABCD^® study(31).

The ABCD^® study encompasses genotype data from 138 plates. 82 individuals from plate 461 were excluded due to potential sequencing issues identified in the genetic data report. Individuals were excluded if their data exhibited a standard deviation of log R ratio >0.35, a standard deviation of B allele frequency >0.1, or an absolute value of wave factor >0.05. These thresholds were chosen to balance data quality with sample retention, informed by both our previous studies(23) and the distributions of these metrics within the ABCD^® study (Supplementary Figure 1). While more stringent (e.g., standard deviation of B allele frequency<0.08) are used in some contexts, the thresholds applied here allow us to retain valuable data while ensuring representation across a broader and more diverse population sample in this study. To ensure identifying reliable CNVs, we used CNVision(36) to merge CNVs detected by two algorithms. Furthermore, we excluded individuals carrying more than 500 merged CNVs, as this threshold may indicate batch effects, genotyping errors, or extreme chromosome abnormalities. This threshold is tailored to the characteristics of specific datasets, and for this population. The distribution of the number of CNVs per individual indicates that the 500 threshold provides the best balance between data quality and sample size (Supplementary Figure 1). CNVs exhibiting >70% overlap between PennCNV(25) and QuantiSNP(26) algorithms were retained for follow-up genetic analysis. All probe coordinates were established based on the GRCh37/hg19 reference genome.

Subsequently, we implemented a series of additional filtering steps aimed at minimizing false discoveries. Specifically, we excluded CNVs smaller than 50,000 base pairs in size, given that smaller CNVs might result from stochastic variations or technical artifacts. This threshold has been widely used in studies investigating CNV associations with human complex traits in large cohorts, including Philadelphia Neurodevelopmental Cohort(23), IMAGEN cohort(33), UK Biobank(37), as well as multi-ancestry cohorts(38). To reduce the variation resulting from noise and increase the true positive rates, we filtered out CNVs spanning <20 SNPs or with PennCNV or QuantiSNP confidence scores <30. These standardized thresholds have been applied in our previous study, resulting in more comparable results across different datasets and studies(23). CNVs demonstrating over 50% reciprocal overlap with segmental duplications, centromeric regions, telomeric regions, or the major histocompatibility complex region were excluded due to the complex genomic structure in these regions. CNVs exhibiting an equivocal number of copies between the two algorithms were also removed. Moreover, in families with multiple members, only one random family member was included in the subsequent analysis.

To describe demographic characteristics of our sample, we extracted age, sex assigned at birth, three measures of socioeconomic status (including parental education, income-to-needs ratio, and area deprivation index) and genetic ancestry estimates (including African, European, East Asian, and American) provided by the ABCD^® study. Parental education measures the highest level of education attained by either parent, ranging from 0 (no formal education) to 21 (PhD). Higher values indicate longer durations of formal education. Income-to-needs ratio reflects household financial circumstances, calculated as the median value of reported household income divided by the U.S. Federal Poverty Guidelines for 2017, adjusted for household size. Ratios below 1 indicate conditions below the poverty level, while values above 1 indicate increasing financial surplus. Area deprivation index quantifies neighborhood socioeconomic conditions, where lower scores suggest wealthier neighborhoods and higher scores indicate more deprived areas. The ABCD^® study offers detailed estimates of genetic ancestry, utilizing genetic principal components analysis to quantify the proportions of an individual’s genetic makeup derived from various ancestral backgrounds. Each reflects the percentage of an individual’s genome that corresponds to African, European, East Asian, and American ancestries, respectively. A higher value in any of these genetic ancestry measures indicates a greater proportion of the individual’s genetic material originates from that specific ancestry. To assess potential biases introduced by the inclusion criteria, we performed a two-sample t-test to compare demographic, socioeconomic and genetic variables between the included and excluded samples. Additionally, a Chi-squared test was performed to examine differences in sex distribution between the two groups.

Psychopathological and cognitive measures

The ABCD^® study includes comprehensive assessments of psychopathological and cognitive measures(39, 40). We derived measures of dimensional psychopathology from the 119-item Child Behavioral Checklist(41) completed for each individual. Parents rated the extent to which specific behaviors (e.g., “Destroys others’ things”) were characteristic of their child over the past 6 months, using a 3-point scale: 0 (“not true”), 1 (“somewhat or sometimes true”), or 2 (“very true or often true”). These 119 items are aggregated to generate raw composite scores for 8 scales: anxious/depressed, withdrawn/depressed, somatic complaints, social problems, thought problems, attention problems, rule-breaking behavior, and aggressive behavior. The sum of scores on these scales provided an overall measure of psychopathology burden, encompassing various behavioral syndromes.

Cognitive performance was assessed by the NIH Toolbox Neurocognitive Battery. The NIH Toolbox encompasses 8 domains(42, 43), including: fluid intelligence (reflecting the ability to solve abstract reasoning problems); dimensional change card sort (measuring ability to plan, organize, and execute goal-directed behaviors); flanker inhibitory control and attention (reflecting the ability to handle multiple environmental stimuli); list sorting working memory (quantifying the ability to store, manipulate, and hold new information); oral reading recognition test (capturing reading ability and academic achievement); picture vocabulary test (characterizing language and verbal intellect); picture sequence memory (representing episodic memory including the acquisition, storage, and retrieval of information); and pattern comparison processing speed (indicating the ability to process new information within a certain amount of time). Scores across these domains were aggregated to generate a composite score reflecting overall cognitive performance.

To reflect phenotypic variability relative to the reference norm, ABCD^® converts overall and each dimensional clinical measures to norm-referenced scores(39). For each cognitive and psychopathological measure, we extracted the standardized score and regressed the effects of age, sex, batch and top 10 genetic principal components by linear regression models. Subsequently, rank-based inverse normalization was applied to ensure the residuals following a normal distribution. In addition, we used two-sample t-test to compare differences in cognitive and psychopathological measures between included and excluded samples.

Genome-wide CNVR association analysis

We utilized ParseCNV2(44) to conduct genome-wide CNVR association analyses for each cognitive and psychopathological measure independently. This method involves mapping individual-level CNV calls to population-level probe-based CNV statistics (separately for deletions and duplications). Subsequently, the significance of each SNP marker was assessed using Fisher’s exact test using linear regression for quantitative (continuous) trait phenotypes. The resulting association statistics were then used to combine SNPs in proximity (default distance of 1MB) with comparable p-values (default threshold of 1 power of 10) into genomic regions referred to as “CNVRs”. Our previous ParseCNV study(45) assessed the validity of this 1M threshold across diverse genomic regions (ranging from 1bp to 2MB) in a dataset of SNP array resolutions similar to the ABCD^® study. The results demonstrated that the 1MB threshold balanced the need to extend CNVRs to incorporate boundary variability while avoiding the overextension that might obscure unique loci. This threshold is also supported by SNP array resolutions that are broadly applicable to variable linkage disequilibrium structures in datasets with diverse genetic ancestries. This CNVR approach offers several advantages over methods that investigate genetic correlates for individual CNVs, increasing power and preventing redundant reporting of overlapping genomic regions associated with the phenotype. Previous studies have successfully applied this approach to identify genomic structural changes associated with brain disorders and physical health(46–48). Additionally, ParseCNV2 identifies the contributing CNVs that drive each CNVR association. To mitigate technical artifacts and reduce the likelihood of false positives, ParseCNV2 retains only recurrent CNVRs that are supported by at least two contributing CNVs, ensuring the associations are not driven by a single individual.

Multiple testing correction of significance

In contrast to genome-wide association scans involving SNPs, a p-value threshold of 5×10⁻⁴ was recommended by the original ParseCNV2 study(44, 45) to define genome-wide significance for CNV (spanning many nucleotides) association testing. This threshold is determined by specific characteristics of CNV analyses, including the number of probes with a nominal frequency of CNV occurrence and the number of CNVRs. Specifically, unlike SNPs, only probes with a nominal frequency of CNV occurrence are informative for association testing, which typically results in fewer than 100 CNV probes across most samples. This substantially reduces the need for the ultra-stringent significance threshold commonly used in SNP-based GWAS (p<5×10⁻⁸). CNVs generally span multiple probes, meaning that probes within the same CNV are not independent observations and should not be treated as such for multiple testing correction. Second, given that most samples containing fewer than 100 detectable CNVRs, a threshold of 5×10⁻⁴ is both statistically justified and sensitive enough to capture meaningful CNV associations without being overly conservative, which would risk overlooking true associations. Finally, considering the correlated structure among 18 psychopathological and cognitive variables (Supplementary Figure 2), we used matrix spectral decomposition to compute eigenvalues for these traits(49, 50). These eigenvalues were subsequently utilized to estimate the effective number of independent tests. In our study, this effective number was estimated to be 11. Consequently, we applied a significance level of p<4.54×10⁻⁵ (i.e., 5×10⁻⁴/11) to correct for multiple testing in the CNVR association analysis across 18 measures.

Longitudinal analysis

The ABCD study recently released 2-year follow-up data for dimensions of child psychopathology, providing an opportunity to assess the longitudinal impact of CNVs. For each CNVR showing association at the genome-wide significance threshold (p<5×10⁻⁴), we identified individuals carrying CNVs contributing to the CNVR association signal. Using both baseline and follow data, we constructed a linear mixed effects model to evaluate the association between CNV status and the corresponding dimension of psychopathology. In this model, CNV status (coded as 1 for individuals carrying contributing CNVs and 0 for those without) and the interaction between age and CNV status were treated as fixed effects. Age, sex, batch, and the top 10 genetic principal components were included as covariates. A random intercept for each subject was included to account for within-subject correlations. The significance of the long-term effect of CNVs on psychopathology development was assessed using the t-value derived from the ‘age-by-CNV status interaction’ factor, with a Bonferroni correction applied at p<0.05 (i.e., 0.05 divided by 11, the effective number of independent tests).

Neurodevelopmental CNVs and diagnosis

In our previous study(22), we compiled a set of recurrent CNVs (48 deletions and 36 duplications) that have been previously associated with genetic risk for neurodevelopmental disorders (hereinafter referred to as neurodevelopmental CNVs). Using this set, we identified individuals in the ABCD study whose CNVs showed at least 50% reciprocal overlap with these neurodevelopmental CNVs. We then constructed linear regression models to separately examine the association between individuals carrying neurodevelopmental CNVs and each psychopathological and cognitive phenotype, while controlling for the effects of age, sex, batch and top 10 genetic principal components. To determine the significance of these associations, we applied a Bonferroni correction at p<0.05 (i.e., p<0.05 divided by 11, the effective number of independent tests). Additionally, the ABCD study collected parent-reported autism diagnoses during enrollment through the question: “Has your child been diagnosed with autism spectrum disorder?” (as indicated by the ‘scrn_asd’ variable). As a complementary analysis, we identified individuals with a parent-reported autism diagnosis and applied the same linear regression models and multiple testing correction to assess their clinical relevance.

CNV annotation and CNV risk scores

The CNV risk score is used to estimate the pathogenic functional consequences of CNV deletions and duplications in individuals -- encompassing genetic burden, intolerance, and dosage sensitivity. CNV burden quantifies the number of genes covered by CNVs or the size of CNVs carried by an individual. Genetic intolerance refers to the annotations of genes within CNVs concerning their ability to withstand loss-of-function mutations. Dosage sensitivity characterizes the sensitivity to the deletion and duplication of CNVs.

We annotated each CNV based on two public databases(51, 52), scoring them in terms of genetic burden, intolerance, and dosage sensitivity (Supplementary Figure 3). Genetic burden was quantified by the size of the CNV and the number of overlapping genes within the CNV. Intolerance metrics included the probability of loss intolerance (pLI)(53) and the loss of function observed/expected upper bound fraction (LOEUF). Specifically, pLI represents the likelihood of a gene being intolerant to loss-of-function mutations, ranging from 0 to 1. A higher pLI score indicates less tolerance to inactivation, implying potentially deleterious functional outcomes. LOEUF compares the observed and expected number of loss-of-function mutations for a gene in a reference population, with values ranging from 0 to 2. Genes with a LOEUF value less than 0.35 are typically considered intolerant. For CNVs, the cumulative inverse LOEUF (iLOEUF) of all encompassed genes is computed. Dosage sensitivity is measured by the probability of haploinsufficiency (pHI) and the probability of triplosensitivity (pTS). These metrics reflect the likelihood of a gene being sensitive to copy number gain or loss, respectively. Thus, the pHI of a deletion is computed as the sum of the pHI values of all genes overlapping with the deletion, while the pTS of a duplication is estimated by the sum of the pTS values of all genes within the duplication. At the individual level, each component of the CNV risk score is calculated as the sum of all deletions and duplications, respectively.

We used a linear regression model to investigate the association of each CNV risk score with individual cognitive and psychopathological measures across 8,564 individuals (the size of the unrelated sample after quality control), while controlling for the effects of age, sex, array batch, and top 10 genetic principal components precomputed by the ABCD^® study. Prior to modeling, rank-based inverse normalization was conducted on each behavioral measure to ensure adherence to a normal distribution. We applied a Bonferroni correction at p<0.05 (i.e., p<0.05/55, calculated as follows: 5 CNV risk scores x 11 (the effective number of independent variables across 18 psychopathological and cognitive phenotypes, see more details in the ‘Multiple testing correction of significance‘ section).

Results

Demographic descriptions and clinical characteristics

After sample quality control, we excluded 3,312 individuals who did not meet the inclusion criteria from the full ABCD^® sample (N=11,876), leaving 8,564 individuals for subsequent analyses (mean age=9.9 years, 4,532 males). Supplementary Table 1 described demographic and clinical characteristics of full sample, included sample, and exclude sample. For demographic variables, we found a higher proportion of males in the included sample compared to the excluded sample (p=9×10⁻⁴). For socioeconomic variables, the included sample had higher parental education (p=0.01) and higher area deprivation index (p=0.001), indicating a higher socioeconomic status and less neighborhood deprivation compared to the excluded sample. Compared to excluded individuals, the included sample showed greater mental health problems in terms of overall psychopathology (p=2.32×10⁻⁵) and five psychopathological dimensions, including anxious/depressed (p=0.03), somatic (8.73×10⁻⁵), thought (p=2.27×10⁻⁶), attention (p=1.62×10⁻⁴), rule-breaking (p=0.02) dimensions. Individuals in the included sample exhibited stronger cognitive performance, particularly in working memory (p=0.01), reading ability (p=0.001), and language function (p=1.93×10⁻⁵). In addition, compared to excluded sample, included sample contained a higher proportion of individuals with African ancestry (p=6.89×10⁻⁴) and a lower proportion of individuals with European ancestry (p=1.64×10⁻⁴).

The CNV landscape of the ABCD^® study

Following conservative quality control (Figure 1), a total of 5760 CNVs were identified across 4111 individuals (Figure 2) while 4453 individuals did not carry any CNVs. Among the detected CNVs, 1391 (24.15%) were deletions (43 homozygous), while 4369 (75.85%) were duplications (161 with copy number >3). While 28.8% of individuals with a CNV carried a single deletion, instances of up to three deletions were observed in nine individuals. Similarly, 61.5% of individuals had one duplication (totaling 2527 out of 4111), with up to four duplications found in 18 individuals (Supplementary Figure 4). Summary statistics at the CNV level were as follows (Supplementary Figure 4): number of SNPs encompassed by deletions, mean=78 (range 20-1811); number of SNPs encompassed by duplications, mean=151 (20–12,635); deletion length, mean=300.99Kb (50Kb-6.14Mb); duplication length, mean=357.76Kb (50Kb-11.10Mb); number of genes within deletions, mean=1 (0-46); number of genes within duplication, mean=2 (0-51) .

Figure 2. The landscape of the CNV calls in the ABCD^® study.

After conservative quality control, we report 5760 CNVs among 8564 individuals from the ABCD^® study. The figure illustrates the chromosomal distribution of all deletions and duplications called in the ABCD^® study, in which blue represents deletions, and red represents duplications.

Genome-wide CNVR association with overall and dimensional psychopathology and cognition

We used ParseCNV2(44) to identify CNVRs associated with variations in each psychopathological and cognitive phenotype. Although no CNVRs reached significance after applying multiple testing correction at p<4.54×10⁻⁵ (i.e., p<5×10⁻⁴ divided by 11, the effective number of independent tests, as detailed in the Methods), we identified 16 CNVRs showing association with psychopathology and cognitive development at a genome-wide significance level of p<5×10⁻⁴ (Figure 3, Supplementary Figures 5 and 6, and Supplementary Tables 2-19). Specifically, for overall psychopathology and cognitive function, we found that a duplication at 10q26.3 (i.e., chr10:133248899-133784881) associated with overall psychopathology (p=3.77×10⁻⁴, standardized beta=3.55, Figure 3A and Supplementary Table 2). Decomposition of this CNVR revealed 16 individuals carrying duplications contributing to the observed association signal. Notably, this genomic region overlaps with two genes, including PPP2R2D and the long non-coding RNA AL450307.1 (Figure 3A). Annotation of this region revealed that common variants within PPP2R2D have been linked to schizophrenia(54), and methylation modifications of the genomic region have been implicated in dimensional psychopathology in youth(55).

Figure 3. Genome-wide CNVRs associations with psychopathology.

(A) Genome-wide CNVR association with overall psychopathology. (B) Genome-wide CNVR association with attention and social problems. Dotted lines in Manhattan plots represent the statistical threshold of p<5×10⁻⁴, the recommended threshold for genome-wide significance in the ParseCNV2 platform. Black lines in Manhattan plots indicate the Bonferroni-corrected p<4.54×10⁻⁵ (i.e., 5×10⁻⁴ divided by 11, which is the effective number of independent tests across 18 psychopathological and cognitive phenotypes, see more details in the Methods). Blue lines under the Manhattan plot represent the CNVs contributing to the CNVR association. The black line and the genes below them indicate the genomic position of genes encompassed by the CNVR associated with psychopathology.

For dimensional psychopathology, the strongest association was observed between attention problems and a duplication at 14q11.2 (i.e., chr14:20247780-20423950, p=7.68×10⁻⁵, standardized beta=2.28, Figure 3B and Supplementary Table 3). Decomposition revealed 717 individuals carrying contributing to this CNVR. This genomic region contains three olfactory receptor family genes, including OR4N2, OR4M1 and OR4K1. Variations in the copy number of olfactory genes have been identified as significantly associated with cognitive decline in patients with Alzheimer’s disease(56). Our CNVR association analysis also identified a duplication at 17q12 (i.e., chr17:34447402-34876195) associated with social problems (p=3.77×10⁻⁴, standardized beta=−3,55, Figure 3C and Supplementary Table 4). This duplication shows a 4.30% reciprocal overlap with a previously reported duplication involving HNF1B that is associated with neurodevelopmental disorders(57). Furthermore, the duplication at chr10:133248899-133784881, which we identified as associated with overall psychopathology, was also linked to somatic complaints (p=1.45×10⁻⁴, standardized beta=3.80, Supplementary Figure 5 and Supplementary Table 5), suggesting somatic problems might be a contributing dimension underlying its association with overall psychopathological variability. In addition, we observed 6 CNVRs associated with other dimensions of psychopathology at a genome-wide significance level (i.e., p<5×10⁻⁴), including chr3:725824-1122130 deletion at 3p26.3 associated with anxious/depressed dimension (p=3.77×10⁻⁴, standardized beta=−3.55), chr15:101999417-102306252 duplication at 15q26.3 associated with withdrawn dimension (p=3.77×10⁻⁴, standardized beta=3.55), chr6:128018507-128090776 duplication at 6q22.33 associated with social problems (p=3.77×10⁻⁴, standardized beta=3.55), chr22:49817455-49907896 deletion and chr6:86708262-87639382 duplication jointly associated with thought problems (p=1.45×10⁻⁴, standardized beta=−3.80 for both), and chr4:148640660-148832349 duplication associated with aggressive behavior (p=1.45×10⁻⁴, standardized beta=3.80, and Supplementary Figure 5 and Supplementary Tables 3-10).

For dimensions of cognitive function, the strongest association was observed between processing speed and a duplication at 4q13.2 (i.e., chr4:69662363-70719533, p=1.47×10⁻⁴, standardized beta=−3.79, Supplementary Figure 6 and Supplementary Table 11). We identified 4 individuals carrying duplications contributing to this CNVR. This genomic region encompasses nine UGT2B family genes, such as UGT2B10, UGT2B7, and UGT2B11. Previous studies(58) and entries in the DECIPHER database(59) have documented individuals carrying this duplication, often presenting with attention deficits, atypical behavior, and intellectual disability. In addition, we identified 5 CNVRs showing marginal associations with other dimensions of cognition (i.e., p<5×10⁻⁴), including the chr22:25653374-25923955 duplication associated with attention performance (p=4.54×10⁻⁴, standardized beta=−1.57), chr21:21787052-21844371 deletion associated with reading ability (p=3.82×10⁻⁴, standardized beta=3.55), chr19:54992016-55102179 duplication associated with processing speed (p=3.82×10⁻⁴, standardized beta=−3.55), chr19:55414173-55535482 duplication associated with language function (p=3.81×10⁻⁴, standardized beta=−3.55), chr9:112298776-112964736 duplication associated with language function (p=3.81×10⁻⁴, standardized beta=3.55, and Supplementary Figure 6 and Supplementary Table 12-19).

Based on longitudinal assessments, we constructed linear mixed effect models to explore developmental trajectories of clinical phenotypes associated with CNVs. Our results revealed that the aggressive behavior dimension was significantly and negatively associated with the interaction between age and a duplication at chr4:148640660-148832349 (t=−6.99, p=2.95×10⁻¹², Supplementary Table 20). This finding suggests that the impact of this CNV on aggressive behavior decreases as age increases, highlighting the dynamic nature of CNV effects over time.

Association of neurodevelopmental CNVs and diagnosis with psychopathology and cognition

To complement the genome-wide scan, we specifically looked for the presence of 48 deletions and 36 duplications that have previously been identified as recurrent CNVs associated with genetic risk for neurodevelopmental disorders (Supplementary Table 21 and Supplementary Figure 7)(22). Among the 4111 individuals from the ABCD^® study, we identified 20 deletions among 30 ABCD^® individuals who showed at least 50% reciprocal overlap(60) with 9 out of 48 neurodevelopmental deletions (Supplementary Table 21). 64 individuals carried 22 duplications showing at least 50% reciprocal overlap with 10 of 36 neurodevelopmental duplications (Supplementary Table 21). Of note, 59 out of 93 individuals carrying neurodevelopmental CNVs had either a deletion at 15q11.2, or a duplication at 15q11.2 or 16p13.11 (Supplementary Table 21). Next, we separately examined whether individuals carrying neurodevelopmental CNVs predicted each psychopathological and cognitive phenotype in the ABCD^® study (see Methods). After applying for Bonferroni correction at p<0.05 (i.e., p<0.05 divided by 11, the effective number of tests across 18 psychopathological and cognitive phenotypes), we found that individuals with neurodevelopmental CNVs had high scores in the social dimension of psychopathology (p=4.20×10⁻³, Figure 4A, Supplementary Figure 8 and Supplementary Table 22). Moreover, these individuals showed poorer performance in terms of overall cognition (p=8.77×10⁻⁴) and three cognitive domains, including fluid intelligence (p=1.57×10⁻⁵), working memory (p=3.59×10⁻⁵), and processing speed (p=2.53×10⁻³, Figure 4B, Supplementary Figure 8 and Supplementary Table 22). In addition, in the ABCD study, parents of a total of 201 children reported an autism diagnosis during enrollment, with 148 children overlapping with the samples we included. We found that individuals with autism were significantly associated with overall psychopathology and all eight dimensions (Bonferroni-corrected p<0.05), with social domain showing the strongest association (p=6.22×10⁻⁴³, Supplementary Table 23). Moreover, these autistic individuals were significantly associated with poor fluid intelligence, attention, working memory (Bonferroni-corrected p<0.05 and Supplementary Table 23).

Figure 4. Associations of psychopathology and cognitive function with individuals carrying known pathogenic recurrent CNVs associated with neurodevelopmental disorders.

(A) Association of overall and dimensional psychopathology with neurodevelopmental CNVs. Individuals with neurodevelopmental CNVs had significantly elevated psychopathology in the social dimension at Bonferroni-corrected p<0.05. (B) Association of global and dimensional cognitive function with neurodevelopmental CNVs. Individuals with neurodevelopmental CNVs had lower performance in term of overall cognition and multiple cognitive domains including fluid intelligence, working memory, and processing speed (Bonferroni-corrected p<0.05). Points in the dot plots indicate the value of a given predictor variable’s standardized effect size and error bars indicate 95%CIs for models of cognition and psychopathology.

CNV risk score association with cognition and psychopathology

At the individual level, we quantified the burden of deletions based on five metrics: the size of deletions (mean=48.89kb [0-6.14Mb]), the number of genes covered (mean=0.13 [0-46]), pLI (mean=0.02 [0-12.77]), iLOEUF (mean=0.17 [0-77.89]) and pHI (mean=0.05 [0-24.27], see Methods and Supplementary Figure 9). For duplications, the CNV risk score at the participant level was determined by the size of duplications (mean=182.51Kb [0-11.10Mb]), the number of genes encompassed (mean=1.26 [0-51]), pLI (mean=0.11 [ranging 0-14.42]), iLOEUF (mean=1.23 [0-86.46]) and pTS (mean=0.295 [0-26.77]). Note that individuals who have no CNVs have CNV risk scores of 0.

To assess the cumulative effect of CNVs on psychopathology and cognition, regression models were performed with each CNV risk score across 8564 individuals, for global and dimensional psychopathology and cognition measures. For deletions, we observed that all domains of the CNV risk score were significantly associated with overall cognition and five cognitive domains at a Bonferroni-corrected p<0.05, including fluid intelligence, attention, working memory, flexible thinking, and processing speed (Figure 5 and Supplementary Tables 24 and 25). Deletion CNV risk scores with the exception of CNV size were also significantly associated with attention problems at a Bonferroni-corrected p<0.05 (Figure 5 and Supplementary Tables 24 and 25). For duplications, greater iLOEUF, pLI, and pTS risk scores were significantly associated with lower overall cognition at a Bonferroni-corrected p<0.05 (Figure 5 and Supplementary Tables 26 and 27). Duplication risk scores were not significantly associated with domains of psychopathology after correction for multiple comparisons (Bonferroni-corrected p>0.05, Figure 5 and Supplementary Tables 26 and 27).

Figure 5. Associations of psychopathology and cognitive function with CNV risk scores.

To determine the cumulative effect of CNVs on psychopathology and cognition, we used regression models to test associations with global and dimensional cognitive functions for each of 5 CNV risk scores across 8564 individuals, including the size of CNV (A), the number of genes overlapping with the CNV (B), the probability of loss intolerance (C), the inverse of the loss of function observed/expected upper bound fraction (D), dosage sensitivity (E). Similarly, we also tested associations with global and dimensions of psychopathology for each of the 5 CNV risk scores (F-J). Points in the dot plots indicate the value of a given predictor variable’s effect size and error bars indicate 95%CIs for models of cognition and psychopathology.

Discussion

The present study delineates the landscape of CNVs in the ABCD^® study and characterizes CNVRs associated with child psychopathology and cognitive function. Individuals in the ABCD^® study carrying recurrent CNVs previously implicated in neurodevelopmental disorders exhibited greater psychopathology in the social domain and lower cognitive performance particularly in the domain of working memory. Aggregated CNV risk scores for deletions were associated with attention domain of psychopathology, as well as with multiple cognitive domains. Additionally, duplication CNV risk scores derived from constraint scores (pLI and iLOEUF) were associated with lower overall cognitive function. These results underscore the significant impact of genomic structural changes on psychopathology and cognitive development in childhood. Furthermore, our study highlights the sensitivity of CNV risk scores as a measure to characterize the complex genetic underpinnings of neurodevelopmental phenotypes.

Previous genetic studies have traditionally focused on identifying genomic structural variants associated with the risk of categorical psychiatric disorders using case-control designs(61, 62). However, case-control designs may not adequately capture the complexity of psychiatric symptoms characterized by high levels of comorbidity and ambiguous diagnostic boundaries(19, 21, 63). In our current study, we sought to characterize CNVRs associated with dimensional psychopathology and cognitive outcomes in childhood. This dimensional approach allowed us to explore how gene dosage imbalance affects dimensional psychopathology in community population that was not ascertained for mental health problems. As compared to pathogenic CNVs discovered in case-control studies, CNVs with a higher frequency in the general population are more likely to be inherited, and may modulate psychopathology and cognitive development without necessarily leading to clinically diagnosed mental illness or intellectual disability.

In our investigation of CNVRs associated with psychopathology measures derived from the Child Behavioral Checklist, we did not find any CNVRs meeting the Bonferroni-corrected significance threshold (i.e., p<4.54×10⁻⁵). However, we identified 10 CNVRs associated with dimensions of child psychopathology at the genome-wide significance level (p<5×10⁻⁴), with the association between the attention dimension and copy number gain at 14q11.2 showing the strongest association signal (Figure 3). Notably, the duplication at 10q26.3 showed associations with both overall psychopathology and somatic dimension, indicating somatic complaints might be a key component contributing to its CNVR association with overall child psychopathology. We also found our identified CNVs showed modest effect sizes in the genetic association analysis. This is likely due to several factors. First, many of the identified CNVs, such as the duplication at 14q11.2 showing marginal association with attention problems (frequency=8.37%), occur at relatively high frequencies in the general population, reflecting normal genetic variation that is not necessarily pathogenic. Given the polygenic nature of cognitive and behavioral traits, individual CNVs contribute only a small proportion to the overall phenotypic variability. Their effects must be considered in the context of broader genetic and environmental influences, which dilute the impact of any single CNV. Second, our study focuses on continuous dimensions of psychopathology, targeting sub-threshold symptoms rather than clinically diagnosable disorders. This approach inherently results in smaller effect sizes compared to studies focusing on extreme phenotypes, such as autism or schizophrenia, where larger CNV effects are often observed. Third, rare CNVs with potentially larger impacts may not have been captured in our analysis due to their low population frequencies and the limited sample size of our cohort, which also reduces power to detect subtle variability in traits. Fourth, the recurrent CNVs included in this study show incomplete penetrance, suggesting some carriers remain unaffected, further contributing to the modest effect sizes observed. Fifth, the “healthy volunteer” selection bias in the ABCD likely underrepresents severely affected individuals. CNV carriers who participated in the study may be higher functioning, while those with more severe impairments might not have been included. This underrepresentation can lead to an underestimation of CNV effects on psychiatric phenotypes. Finally, a related challenge is that the Child Behavioral Checklist, which is based on parent report of items on a 3 point scale, provides a relatively limited depth of phenotypic characterization(41). More sensitive phenotypic characterization by aggregating measures across future ABCD^® time points may yield increased power to identify CNVRs associated with developmental psychopathology.

To better understand the contributions of common and rare genetic variants to psychopathology, we extracted SNP association signals from meta-analyzed genome-wide association scan across 12 psychiatric disorders(64). Based on the colocalization observation, we found common genetic variants within CNVRs associated with overall psychopathology showed weak association with psychiatric disorders, with the strongest association signal occurred at the rs10030847 (p=0.0003). In the context of previous evidence of enriched genome-wide association signal within pathogenic CNVs(33), this observation has several potential explanations. First, SNP-based association results were primarily derived from cohorts with genetically-defined European ancestry, In contrast, our CNV analysis was conducted in a multi-ancestry cohort, which introduces confounding effects related to ancestry when attempting to model common and rare genetic variants together. Second, it is possible that SNPs tag a genetic haplotype with an inherited CNV, such that the GWAS SNP effect is capturing the effect of this CNV. Third, common genetic polymorphisms with subtle effects, acting as genetic modifiers of complex phenotypes, could act cumulatively with CNVs and environmental factors to modulate variability in psychopathology and cognitive ability during childhood. Fourth, both CNVs and SNPs may influence phenotypes through regulatory effects on nonadjacent genomic regions, which complicates the interpretation of simple overlap between CNVRs and GWAS summary statistics. Consequently, the genetic liability for adult psychiatric disorders may not fully reflect the genetic architecture underlying subclinical psychiatric symptoms during adolescence. It is important for future work to model CNV and SNP effects together in order to unpack any overlap in genetic risk factors. Previous work has suggested additive effects between CNV risk scores and polygenic risk scores(23), as expected given that pathogenic CNVs are disproportionately de novo variants.

After applying a stringent Bonferroni correction for multiple testing, no genomic regions were significantly associated with cognitive measures. However, six CNVRs demonstrated associations at the genome-wide significance level (p<5×10⁻⁴), offering cytogenetic insights into deviations in cognitive development trajectories in children. For example, poor performance in processing speed was associated with a duplication at 4q13.2, which encompasses several genes. Individuals carrying this duplication often present with attention deficits, atypical behavior, and intellectual disability, as reported in clinical studies(58) and the DECIPHER platform(59). Additionally, a duplication at 22q11.23, which includes LRP5L, was associated with the attention domain. This CNV has previously been reported in individuals with juvenile myoclonic epilepsy, a condition characterized by delayed cognitive development. On chromosome 19, duplications at 19q13.42 were associated with poor processing speed and language ability. Pediatric clinical studies have suggested long-term follow-up for children with 19q13.42 microduplications, as these duplications may have subtle but lasting impacts on cognitive development, despite no observable prenatal or postnatal growth delays(65). Our findings provide additional evidence for the potential protracted effects of this CNV on cognitive function, as revealed through a genome-wide CNV screening. It is worth noting that decomposition analysis identified relatively few individuals carrying CNVs contributing to these marginal associations. This is likely due to the modest sample size analyzed in this study, which limits statistical power and the ability to detect more genomic regions associated with clinical phenotypes. Nevertheless, ParseCNV2 ensures robustness by defining CNVRs based on recurrent CNVs (i.e., involving at least two individuals), thereby reducing the likelihood of false positives and technical artifacts. Future studies with larger cohorts will be critical to validating these findings and uncovering additional CNVs associated with cognitive development.

Extending our analysis to other known recurrent, pathogenic CNVs associated with neurodevelopmental disorders(22), we found that individuals carrying neurodevelopmental CNVs exhibited lower performance across almost all cognitive domains, a finding that is consistent with recent adult studies(66). However, with the exception of deficits in the social domain, individuals carrying neurodevelopmental CNVs did not have significant differences in measures of psychopathology (Figure 4). The lack of more consistent effects on psychopathology may be indicative of heterogeneity in psychiatric presentation of neurodevelopmental CNVs, as compared to relative homogeneity of the cognitive impact across domains. It is also possible that the cognitive impact is more apparent during childhood (or more readily measured with structured batteries). Some individuals with current cognitive deficits may eventually progress to clinically diagnosed psychiatric disorders, as additional psychiatric manifestations may unfold on a developmental timeline across adolescence and adulthood, depending on interactions with environmental and other genetic factors (e.g., risk due to common genetic variants).

Compared to identifying specific CNVRs or querying known recurrent CNVs, CNV risk scores offer a complementary approach based on the expected functional consequences aggregated across CNVs, particularly in terms of encompassed genes’ tolerance to loss-of-function mutations and dosage sensitivity(22, 23). In our analysis of the association between CNV risk scores and psychopathology, we observed that deletion risk scores were associated with variations in attention domains. However, similar associations were not observed with any duplication risk scores, despite duplications at 10q26.3 and 14q11.2 being marginally associated with psychiatric symptoms in the genome-wide CNVR association analysis. One potential explanation for this discrepancy is that the effect sizes of duplications is generally lower and also more heterogenous compared to the effect sizes of deletions(67, 68). Interestingly, unlike other components of CNV risk scores, we found that the size of deletions did not significantly correlate with any psychopathological dimensions. This underscores the potential for improved CNV risk scores, possibly attuned to specific phenotypic associations, to be explored in future investigations.

Our study identified CNVRs marginally associated with psychopathology and cognitive traits in children, which advances our understanding of genetic influences on these dimensions. However, we believe it is premature to advocate for clinical reporting of these CNVs due to their current lack of validation as clinical markers. Reporting such variants could lead to misinterpretation, as these CNVs do not predict specific disorders but are associated with a spectrum of behavioral or cognitive variability. Ethically, reporting could also risk stigmatization or undue concern, especially in cases where families might misunderstand these findings as predictors of future impairment. Nonetheless, combining these findings with family genetic information from high-risk groups could offer new insights into the familial and developmental effects of inherited or de novo mutations on early-onset psychopathology. Ongoing research and validation will be essential to responsibly integrate these findings into clinical frameworks.

Several methodological limitations should be noted in the present study. First, our CNV association results were constrained by the limited sample size. Small samples are more susceptible to false discoveries in genetic analysis(69), particularly with rare genetic variants that may occur only once in a small population. However, given our strict quality control and the fact that the CNVR with the smallest sample size recovered a known pathogenic CNV, false discovery due to low sample size is unlikely to be an issue for the present study. More relevantly, small samples may overlook genetic loci influenced by the nuanced variability of phenotypes among individuals(70). It is essential to collect larger child cohorts to replicate the current genetic findings, explore additional genomic structural variants related to cross-diagnostic psychiatric symptoms, and investigate how genetic events interact with environmental factors to modulate mental and cognitive development. Second, a technical limitation concerns the genotype platform used for inferring CNVs. Although this approach is generally considered reliable, it lacks SNP probes for detecting CNVs in repeat-enriched regions due to the complex properties of genomic architecture(34, 71). Employing advanced whole-genome sequencing platforms could enhance the detection of high-fidelity CNVs(72, 73). However, genotype array platforms remain the most cost-efficient option for calling reliable CNVs at the current stage. Third, while ABCD^® used epidemiological sample strategies, the sample is not truly “population representative” and is likely to skew towards higher-functioning children. Consequently, children with severely affected CNVs were less likely to participate in the ABCD^® study, leading to a decrease in power to detect CNVRs in the present study (e.g., compared to a clinically-ascertained sample) and an underrepresentation of the contribution of CNVs to neurodevelopmental phenotypes.

Cumulatively, the present results represent a significant advance in our understanding of CNVs in the ABCD^® study. In sharing detected CNVs and analysis code for use by other researchers, this resource will be valuable in future work aimed at understanding CNVs’ impact on interindividual variability in complex traits including neurocognitive development and child psychopathology.

Data Availability

CNV calls will be made available on NDAR (https://nda.nih.gov/study.html?id=2589), and code to derive CNVs and perform quality control is publicly available on Github (https://github.com/BGDlab).