Introduction

Behavioral disinhibition (BD) is a latent quantitative measure designed to capture common variation shared across many harmful or dangerous behaviors including substance problems, antisocial or criminal behavior, and novelty seeking (Young et al. 2000). In addition, 60–80 % of variation in BD is attributed to additive genetic effects, making BD more heritable than many of the individual component behaviors used to formulate the latent BD construct (Young et al. 2000; Krueger et al. 2002; Hicks et al. 2013). To date, genome-wide association studies (GWAS) have been restricted to such individual component behaviors of BD (McGue et al. 2013), including use or abuse of various substances (e.g. alcohol (Bierut et al. 2010; Edenberg et al. 2010; Frank et al. 2012; Gelernter et al. 2014a; Kapoor et al. 2013; Schumann et al. 2011; Treutlein et al. 2009; Wang et al. 2012), tobacco (Bierut et al. 2007; Tobacco and Genetics Consortium 2010; Liu et al. 2010a, b; Thorgeirsson et al. 2010), cannabis (Agrawal et al. 2011; Verweij et al. 2013), methamphetamine (Uhl et al. 2008), opioids (Gelernter et al. 2014b; Nielsen et al. 2010), and cocaine (Gelernter et al. 2014c)), conduct disorder (Dick et al. 2011), adult antisocial behavior (Tielbeek et al. 2012), and related personality constructs such as excitement seeking (Terracciano et al. 2011). Although certain behaviors, most notably tobacco use (Bierut et al. 2007; Tobacco and Genetics Consortium 2010; Liu et al. 2010a, b; Thorgeirsson et al. 2010), have identified robust associations with specific variants, many GWAS fail to identify individual loci with genome-wide significant effects. This suggests that much of the heritability underlying each trait is unlikely the result of a small number of variants with large effects, and will require larger sample sizes in order to identify variants with small effects (Lee et al. 2011). GWAS of other phenotypes have identified significant replicated effects when large enough samples sizes have been amassed to provide adequate statistical power to identify variants despite very small effect sizes (e.g., accounting for 0.1 % of the total variance, or less; Sullivan 2011; Rietveld et al. 2013).

Increasing sample sizes is only one of a number of ways to increase statistical power. Improved phenotypic assessment and modeling could also provide increased statistical power for studies conducted in more moderately sized samples, and particularly for phenotypes that are presumed to be continuously distributed in the population (van der Sluis et al. 2013). BD is a prime example in this context, as relevant quantitative differences in phenotypic severity are maintained between individuals, whereas a case–control approach is fairly insensitive to these differences. However, one issue with searching for specific genetic influences on many continuous phenotypes, such as BD, is that the most severe, clinically significant levels are relatively rare in the general population, as they are located on the extreme ends of the distribution. Ascertaining samples specifically for individuals with extreme phenotypes may improve our ability to detect small genetic effects by increasing the sample variance. Therefore, an ideal sample might be considered one that is enriched (and well-measured) for extreme BD characteristics.

We report here results and characterization of the initial GWAS from the Center on Antisocial Drug Dependence (CADD), an adolescent sample over-selected for severe BD characteristics. Any genetic effects on BD are potentially attributable to many (i.e., thousands of) variants, each with a very, very small effect. Incorporating methods for aggregating effects across multiple variants, such as gene- and pathway-based analyses, can identify promising causal biological systems beyond the significance of any single variant. In addition to SNP level association, the current study applied gene-based, pathway-based, and genome-wide approaches to characterize genetic influences on BD in a diverse, clinically-oversampled, thoroughly phenotyped sample. By supplementing a GWAS with several methods of aggregating genetic evidence across many potentially associated variants, we sought to generate novel insights into the potential genetic etiology of BD and identify promising candidates, either old or new, for future study.

Methods

Participants

Participants with genetic and relevant phenotypic data were ascertained from the CADD projects; full details of participant selection for inclusion in the GWAS sample are provided in the Supplemental materials. GWAS participants were drawn from several primary studies described elsewhere (Hartman et al. 2008, Petrill et al. 2003, Rhea et al. 2006; Stallings et al. 2005). The current sample of 1,901 unrelated adolescents was over-selected for adolescent BD, with half of the participants ascertained specifically from high-risk populations (i.e. recruited through substance abuse treatment, special schools, or involvement with the criminal justice system; see Supplement for additional criteria for clinical probands). CADD GWAS participants had an average age of 16.5 (SD = 1.4, range = 13.0–19.9), 28.9 % were female, and 37.3 % of participants reported non-Caucasian ancestry (primarily Hispanic or African; see Supplemental Table S1 for complete demographic statistics).

Phenotype

BD was defined as a composite measure of substance dependence vulnerability (assessed across 10 substances), novelty seeking, and conduct disorder symptoms. The BD phenotype has been previously examined within the CADD samples, including Young et al. (2000) demonstrating that the component measures have loadings ≥0.4 on a common, highly heritable BD latent factor, and linkage analyses by Stallings et al. (2003, 2005). A full description of construction of the BD phenotype is provided in the Supplement; Supplemental Fig. S1 shows the distribution of BD in the CADD GWAS sample. Briefly, principal component scores were normed to community-representative samples in CADD and applied to all CADD GWAS participants from both the community-representative (48.2 %) and high-risk samples (51.8 %). Average scores on the BD composite measure were 0.19 (SD = 1.2, range = −1.9 to 5.0) for the community-representative participants and 2.76 (SD = 1.2, range = −0.3 to 6.7) for the high-risk participants.

Genotyping

All participants were genotyped on the Affymetrix 6.0 platform (Affymetrix, Inc., Santa Clara CA), with a total of 696,388 autosomal SNPs available for analysis after quality control. Full details on processing and cleaning genotypes for the CADD GWAS sample is provided in the Supplement. Population stratification was examined by performing multidimensional scaling in PLINK (Purcell et al. 2007), in which ten ancestry dimensions were estimated. The first three dimensions notably captured genetic variation among individuals of self-reported African, Hispanic, and Asian ancestry, compared to a central (majority) node of individuals of self-reported European ancestry. Supplemental Fig. S2 illustrates the first three ancestry dimensions within the CADD GWAS sample (along with individuals’ self-reported ancestry).

Analyses

Genome-wide analysis was conducted as a linear regression of the additive effect of each SNP on BD in PLINK (Purcell et al. 2007). All autosomal SNPs that passed basic quality controls were tested for association with BD, and 10 ancestry dimensions were included as covariates. Age and sex were accounted for in the estimation of the BD phenotype. The criterion for individual SNP significance was set at p < 5 × 10−8.

Genome-wide data from the CADD GWAS sample were further characterized using genome-wide complex trait analysis (GCTA; Yang et al. 2011). GCTA allows us to estimate the proportion of variance in the phenotype that may be explained using all of the genotyped SNPs using restricted maximum likelihood (REML) analysis. While this method does not specifically identify any causal variants, it does estimate the total proportion of sample variance that may be explained by all of the genotyped SNPs.

Gene-based tests were conducted using VEGAS (Liu et al. 2010a, b), which aggregates evidence of association across all SNPs within a gene. A total of 16,094 autosomal genes were tested for association with BD in CADD, based on the primary GWAS results, with a multiple-testing-corrected significance threshold set at p < 3.1 × 10−6.

INRICH (Lee et al. 2012) was selected to conduct our pathway analyses as it is well-suited for testing both large (i.e., exploratory) and small (i.e., candidate) pathway sets. We took two, complementary approaches to pathway analysis: first, we sought to confirm previously proposed candidate gene pathways (Hodgkinson et al. 2008); second, we conducted an exploratory analysis aimed at identifying novel pathways involved in BD (The Gene Ontology Consortium 2000). Additional details of the pathway analysis methods are discussed in the Supplemental Materials.

Promising results from the pathway analysis of the CADD sample were followed up in two additional samples: the Minnesota Center for Twin and Family Research (MCTFR; N = 3,378), a community-based adolescent sample (McGue et al. 2013; Miller et al. 2012), and the Study of Addiction: Genes and Environment (SAGE; N = 3,988), a clinically over-selected study of addiction (Bierut et al. 2010; dbGaP study accession: phs000092.v1.p1). A phenotype similar to BD as defined in the CADD sample was available in the MCTFR sample (Hicks et al. 2011; McGue et al. 2013). The phenotype analysed in the SAGE sample was the average number of dependence symptoms for substances that each participant used. Full description of the MCTFR and SAGE samples is provided in the Supplement.

Results

Figure 1 summarizes the GWAS results for BD in the over-selected CADD sample. No individual SNP reached genome-wide significance (p < 5 × 10−8), nor did any SNP reach genome-wide significance in the MCFTR or SAGE samples (see Supplemental Fig. S3 for QQ plots of the GWAS results from each study). Results from loci reaching p < 5.0 × 10−5 in CADD are summarized in Table 1 (full GWAS results are available from the first author on request). The most significant SNP in the CADD GWAS was rs7104461 (p = 5.8 × 10−6), an intergenic SNP on chromosome 11 for which there are no previously reported associated phenotypes. While this SNP was not genotyped in either the MCTFR or SAGE samples, it is in linkage disequilibium with rs341058 (r 2 = 1.0 in 1,000 Genomes Pilot 1 CEU sample, distance = 8,721 bp; Johnson et al. 2008), which was genotyped on both MCTFR and SAGE platforms and may serve as a proxy to compare results across samples. This proxy SNP was not associated with either adolescent BD in MCTFR (p = 0.30) or adult substance dependence symptoms in SAGE (p = 0.87).

Fig. 1
figure 1

Plot of −log10(p) from the CADD GWAS, arranged by chromosomal location. The top (dashed) horizontal line indicates genome-wide significance at p = 5 × 10−8; the lower (solid) line marks p = 5 × 10−5 (loci described in Table 1)

Full size image
Table 1 Top associated loci from the CADD GWAS
Full size table

Whole-genome SNP-heritability was estimated with GCTA in the CADD sample. SNPs genotyped in the current study explained 27.8 % of the CADD sample variance in BD (SE = 0.15, p = 0.03). The point estimate of heritability remained fairly stable when the sample was restricted to individuals estimated to be <2.5 % identical-by-state (N = 1148, V(G)/Vp = 30.9 %, SE = 0.28, p = 0.10) or those individuals with only Caucasian ancestry (as determined by an examination of ancestry component plots, N = 1031, V(G)/Vp = 49.3 %, SE = 0.31, p = 0.06).

Gene-based association tests identified seven genes as significant after Bonferroni correction for testing >16,000 genes: MAGI2 (p < 1.0 × 10−6), NAV2 (p < 1.0 × 10−6), CACNA1C (p = 1.0 × 10−6), PCDH9 (p = 1.0 × 10−6), MYO16 (p = 1.0 × 10−6), IQCH (p = 2.0 × 10−6), DLGAP1 (p < 1.0 × 10−6). We examined overlap of these novel “candidate” genes derived from the CADD GWAS with results from MCTFR and SAGE as a single “pathway” (i.e., gene set) in INRICH (Lee et al. 2012). This allowed us to estimate whether specific genes identified in the CADD results overlapped with the low p value genomic regions (i.e., loci tagged at r 2 > 0.5 by a SNP reaching GWAS p < 5 × 10−3) in the MCTFR and SAGE results more than expected by chance. The CADD-identified gene set was not significant in analysis of either the MCTFR (0 regions overlapped genes identified in CADD, p = 1.0) or SAGE samples (6 regions overlapped genes identified in CADD, p = 0.14).

Supplemental Table S2 presents gene-based association test results for previously identified addiction candidate genes (Hodgkinson et al. 2008), none of which were significant after adjustment for multiple testing (minimum p = 1.4 × 10−3). Supplemental Table S3 gives results for each of the addiction candidate gene sets tested in CADD. None of the addiction candidate gene sets showed evidence of greater-than-chance overlap with low p value genomic regions in the CADD GWAS (minimum p = 5.0 × 10−1).

Promising pathways emerging from our exploratory pathway analysis were defined as those meeting nominal significance before correcting for multiple testing in CADD and either MCTFR or SAGE samples (Empirical p < 0.05). Two pathways met these criteria: visual perception (Empirical p CADD  = 0.038, p MCTFR  = 0.012, p SAGE  = 0.22) and phosphatidylcholine biosynthetic process (Empirical p CADD  = 0.039, p MCTFR  = 1.0, p SAGE  = 0.026). Neither pathway achieved marginal significance in any sample after correction for multiple testing (i.e., Corrected p < 0.10). Supplemental Table S4 provides results from all 72 pathways meeting Empirical p < 0.05 in CADD (from a total of 3,440 pathways tested) that were subsequently tested in the MCTFR and SAGE samples.

Discussion

No SNP was significantly associated with BD in the CADD GWAS. This is not surprising, given the relatively small sample. GWAS of psychiatric and behavioral phenotypes that have successfully identified and replicated individual effects of common SNPs have relied on very large samples (Rietveld et al. 2013; Ripke et al. 2013). Despite the lack of significance of any individual SNP, GCTA REML analysis estimated that 49.3 % (SE = 0.31, p = 0.06) of the Caucasian ancestry sub-sample variation in BD could be accounted for by all of the genotyped SNPs. Conversely, a similar study found no evidence of variance in early adolescent (12-year-old) non-substance behavioral problems being attributable to common variants (Trzaskowski et al. 2013). This may suggest qualitative differences between genetic effects on BD at different ages, an effect that has been reported from twin models of comorbidity between dependence on different substances (Vrieze et al. 2012), which is a marker of BD.

Gene-based tests identified seven genes associated with BD in the CADD sample. However, neither the genes nor pathways identified as marginally overrepresented in the CADD GWAS results showed evidence of replicable low-p values in either the MCTFR or SAGE samples. Taken together, these findings suggest that discoverable effects of common SNPs underlie the genetic architecture of BD, although better-powered studies are required to identify the associated loci.

The comparisons made between datasets must be considered in light of several limitations of the current study. There are substantial differences among the examined samples in terms of age (CADD and MCTFR represent adolescent data, while SAGE was comprised of adults), sex composition (MCTFR and SAGE are split evenly by sex, while CADD has an overrepresentation of males due to the sampling scheme), and diversity of ancestry (MCTFR is less diverse than either CADD or SAGE, which each have different representations of non-Caucasian ancestry groups). The sampling schemes of CADD and SAGE aimed to increase power to detect effects by oversampling extreme phenotype individuals, whereas the MCTFR study is closer to community-representative.

We sought to identify genetic influences on adolescent BD through a multifaceted approach. We initially characterized results from a standard GWAS by estimating the variance explained by common SNPs, and used gene- and pathway-based tests to identify potential novel candidate genes and pathways. Results from the estimation of sample variance explained by all genotyped SNPs and significant gene-based tests suggest there is a real genetic signal to be detected within the noise. However, the current sample is likely underpowered to detect realistic effect sizes of individual SNPs. Further, the lack of correspondence between pathway analyses in the CADD and replication samples may be due to limited power, or qualitative differences in the genetic effects on BD across different ages (adolescent versus adult) or sampling distributions (over-sampled for BD versus community-representative). Key to the search for causal genetic pathways underlying BD will be the availability of increasingly large, thoroughly phenotyped samples. Although the current analyses did not identify specific loci associated with BD, we demonstrate substantial heritability due to effects of common SNPs. Larger studies with appropriate phenotypes could well allow successful identification of common variants associated with BD.