The genetic architecture of pediatric cognitive abilities in the Philadelphia Neurodevelopmental Cohort

The Philadelphia Neurodevelopmental Cohort

The Philadelphia Neurodevelopmental Cohort (PNC) is a population-based sample from the greater Philadelphia area, including over 9000 individuals ages 8-21 years who received medical care within the Children's Hospital of Philadelphia (CHOP) network. The participants presented for a diverse set of medical needs, ranging from a general health checkup and minor problems (e.g. sports related bruise, rash) to chronic condition management (e.g. asthma, type 1 diabetes) to potentially life threatening health problems (e.g. cancer). They were initially enrolled in the genetic study at the Center for Applied Genomics (CAG) at CHOP. Upon assent/consent, the participants were genotyped during the time of their clinical visit and provided written permission to be recontacted for studies of complex pediatric disorders. The PNC participants were selected at random after stratification by sex, age and ethnicity. The overall inclusion criteria included: 1) ability to provide signed informed consent (parental consent was required for participants under age 18), 2) English language proficiency, and 3) physical and cognitive ability to participate in computerized cognitive testing.

All PNC participants completed the Computerized Neurocognitive Battery (CNB) and were assessed psychiatrically with a structured interview. The CNB consists of tests that have been used in functional neuroimaging to probe specific brain systems and is administered with a web browser. It assesses performance on a range of cognitive tasks. The CNB was designed to capture variation in four ability domains, and includes three specific tasks within each domain: 1) executive control (abstraction and mental flexibility, attention, working memory); 2) episodic memory (verbal, facial, spatial); 3) reasoning (verbal, nonverbal, spatial); and 4) social cognition (emotion identification, emotion differentiation, age differentiation). The specific measurement strategy employed for each of the 12 tasks has been described elsewhere (20), but a summary of the measures and their psychometric properties in the PNC is included in the supplemental material (Table S1). The battery also included the reading items from the Wide Range Achievement Test (WRAT) (21).

Cleaning and imputation of genotype data

This study employed genome wide complex trait analysis (GCTA) to estimate the fractional contribution of common SNPs to phenotypic variation in cognitive ability in the general population. One can reduce bias in values estimated through GCTA by minimizing ancestral heterogeneity in the sample (8, 22). As the PNC cohort was drawn from a diverse United States urban population, these analyses were limited at the outset to the subset of participants who identified themselves as white non-hispanic (WNH; n=5,141). All samples were genotyped on one of three Illumina arrays: the HumanHap550, HumanHap610, or OmniExpress v2. Within the self-described WNH group, population outliers were further excluded based on directly genotyped SNP data, prior to imputation. Data were cleaned using a standard approach (23)(Supplement), which reduced the sample by 584 individuals. Over half (62%) of these individuals were excluded for excess relatedness (the PNC included siblings). We conducted a principal components analysis in PLINK (24) (Figures S1a-S1d) which identified 527 individuals with outlying ethnicity, who were subsequently removed. An additional 341 individuals were removed in further phenotypic and genotypic exclusions, described below, resulting in a final analytic sample of 3,689 individuals.

The genotype data were imputed in a separate phase of the study at CHOP. Unobserved genotypes from each chip set were imputed using the IMPUTE2 package and the reference haplotypes in Phase I of the 1000 genomes data (June 2011 release) that included approximately 37,138,905 variants from 1,094 individuals from Africa, Asia, Europe and the Americas. Methodological details regarding the imputation are provided in the Supplement. The imputed genotype data were used in the GCTA analyses.

Phenotypic analyses

We created an index of the overall severity of participant medical conditions based on information obtained from parent or adult proband (ages 18-21) interviews and electronic medical records (EMR). The index included the following levels: 1) None or Minor (43.4% of analytic sample)- no ongoing medical conditions requiring sustained intervention (e.g. well child visit or sprained ankle); 2) Mild (33.5% of analytic sample)- conditions requiring pediatric visits and at times medications but mild in severity (e.g. asthma, allergies); 3) Moderate (22.8% of analytic sample)-medical conditions requiring standing medications and monitoring (e.g. diabetes; lupus); 4) Severe- medical conditions requiring multiple procedures and monitoring that can be life threatening. Individuals with medical rating 4 (n=274) were excluded from the analysis to avoid conflation of their cognitive ability with the physical symptoms of their medical phenotype, which may influence test-taking performance.

Age of the participants was distributed uniformly between 8 and 21 years (mean=13.7 years). Both age and sex were associated with mean differences in CNB performance (20), and were regressed out of all CNB variables, as well as the WRAT, prior to analysis. Individuals with scores more than four standard deviations from the mean of any CNB variable were designated missing for that variable. Using the age- and sex-regressed scores, we conducted common factor and principal components analyses to examine their correlation structure. Common factor analysis (CFA) analyzes the structure of the shared variance of a variable set; principal components analysis (PCA) examines all observed variance. We generated common factor and principal components scores—continuous variables that indicate individual rank within the standardized distribution of the factor or principal component—for use in the genetic analyses. Both the CFA and PCA analyses used promax rotation, allowing the factors to be correlated. Phenotypic analyses were conducted in SAS version 9.3 (25). An unrotated principal components analysis was conducted for purposes of comparison and yielded similar results (Table S2). Given the strong correlation between the variables, only the rotated factor and principal components scores were included in the genetic analyses. There were 46 individuals missing the cognitive battery who were not included in the analysis.

GCTA

The approach underlying GCTA has been described previously (5, 8, 22). In brief, one uses common SNPs to create a genomic relatedness matrix (GRM) that includes pairwise relatedness estimates for all individuals in a sample. The GRM can be estimated using all typed SNPs (across the genome) or using a particular subset of SNPs, for example those on a specific chromosome. Using the GCTA software tool (8), one compares phenotypic similarity with genotypic similarity, which provides an estimate of the fraction of phenotypic variation attributable to the additive effects of common SNPs (SNP-based heritability). For example, a univariate GCTA estimate of 0.35 suggests that 35% of variation in a trait is attributable to the SNPs used to make the GRM, as well as additional variants with which the typed markers are in linkage disequilibrium (LD; correlated variants). If the GRM was estimated using all genotyped SNPs, that fraction reflects the contribution of all common SNPs; if the GRM was estimated using just the SNPs on a single chromosome, that fraction reflects the specific contribution of common SNPs on that chromosome.

We constructed the primary GRM from all imputed, autosomal SNPs with minor allele frequency greater than 1% and Impute2 info score, an imputation quality metric, greater than 0.6 (n=7,635,695 SNPs) (26). This GRM was used in the primary GCTA analyses. After constructing the GRM, we removed the remaining individuals with relatedness above 0.05 (n=20) (26), which yielded the final analytic sample of 3689 individuals. In the univariate analyses, we first estimated the fraction of phenotypic variation in each cognitive trait explained by all common SNPs. Secondly, we estimated genetic correlations between each of the traits with nominally significant (p<0.05) univariate values. Genetic correlations provide an estimate of the degree to which the common variant influences on two traits are shared. A correlation of 1 suggests that the traits share all of their common genetic influences; a correlation of 0 suggests their influences are independent.

The second set of analyses examined the genomic distribution of genetic influence on the accuracy factor scores. We segmented the 7.6 million SNPs used in the primary GRM into: a) 22 chromosome-specific GRMs and b) genic or intergenic space using three genic window definitions(22). For each gene, all genic footprints were defined using the RefSeq transcript with the longest spliced protein-coding DNA sequence (27). All genome files were adjusted for incomplete LD between tagged and causal SNPs, as described by Yang et al. (2011). All analyses controlled for the first ten principal components of ancestry, which collectively explained less than 2% of phenotypic variation in the common factor score.

Phenotypic Analyses

The common factor analysis suggested that the measures have a single factor solution; only one factor had an Eigenvalue exceeding one (Figures S2a). In PCA, accuracy had a dominant first component but two additional interpretable components with Eigenvalues greater than one. The first principal component consisted of the reasoning and executive function items, along with the WRAT. The second principal component included the social cognition items; the third principal component included the memory items (Figure S2b). Factor loadings for both the full sample and age-specific subgroups, as well as the phenotypic correlations, are included in the Supplement (Table S3a-b).

Univariate GCTA Analysis

Table 1 presents the univariate GCTA results. The individual trait scores suggested a wide range of common variant influence, ranging from 0% (emotion differentiation) to 43% (WRAT) of phenotypic variation explained. The reasoning traits had the strongest SNP-based heritabilities, with estimates between 0.30 (s.e.= 0.10; p=2e-04) and 0.41 (s.e.= 0.10; p=4e-06). The heritability estimates of the factor and principal component scores reflected the results of their component traits—the first principal component score (reasoning/executive function) was the most genetically influenced of the summary measures (0.46; s.e.= 0.11; p=8e-06). Strong common variant influence was also suggested for the common factor score (0.36; s.e.=0.11; p=5e-04). As indicated in Table 1, GCTA estimates for 7 of the outcome variables remained significantly different from 0 following a Bonferroni correction for 17 independent tests, which is conservative given that the outcome variables are correlated.

Bivariate GCTA Analysis

First, we estimated bivariate correlations between each of the variables with nominally significant (p<0.05) univariate heritability (Table 2). The genetic correlations ranged from -0.25 (emotion identification and nonverbal reasoning; s.e. 0.20; p(r_g=0)=0.09) to 1.00 (WRAT and language reasoning; s.e. 0.15; p(r_g=0)=2e-07). More than forty percent of variation in the genetic correlations (40.8%) was explained by the phenotypic correlations (p=0.01), which led the variables to cluster genetically in a manner similar to that predicted by the phenotypic factor and principal component analyses. As shown in Figure S3, the genetic correlations were, on average, greater than the phenotypic correlations, reflecting substantial pleiotropy across cognitive traits.

Second, we correlated the individual tasks (e.g. language reasoning) with the common factor score. These correlations indicate the strength of the genetic relationship between the variable and the factor, and the extent of genetic influences on the variable that are distinct from those on the factor (Table S4). With the exception of the association between emotion identification and the factor score (r_g=0.18; s.e. 0.18; p(r_g=0)=0.2), each of the genetic correlations exceeded 0.5. However, of the six item-to-factor genetic correlations, four were significantly different from one (p(r_g=1)<0.05). This further suggests that while many of the genetic influences on cognitive ability are shared, cognitive domains often also have unique genetic influences.

Genomic partitioning

We estimated the association between chromosome length and phenotypic variation explained per chromosome for the common factor score, which suggested that the factor is highly polygenic. Chromosome length explained 31.1% (p=0.007) of variation in GRM prediction (Figure 1). The polygenic model is further supported by the regional analyses. SNPs in both genic and intergenic space were significantly associated with variation in cognitive ability (p(joint association with genic and intergenic regions)=0.009-0.08; Figure S4).