Non-Hispanic European–American adults who completed the Study of Addiction: Genetics and Environment (SAGE; Bierut et al., 2010) were included in analyses (N = 2573; see Table 1 for demographic information; see Supplementary Materials and Methods for details regarding ancestry determination). Alcohol dependent (n = 1160; required 12-month clustering of DSM-IV symptoms) and control (n = 1413) participants were recruited from three large, complementary datasets ascertained for alcohol (Collaborative Study of the Genetics of Alcoholism; Reich et al., 1998; Foroud et al., 2000), nicotine (Collaborative Study of the Genetics of Nicotine Dependence; Bierut et al., 2007; Saccone et al., 2007), and cocaine (Family Study of Cocaine Dependence; Bierut et al., 2008) dependence. Alcohol dependent cases often met criteria for a variety of other substance use disorders. Controls did not meet criteria for alcohol dependence or for cocaine, cannabis, and opioid dependence (nicotine dependence was allowed) but may have used these substances and endorsed some symptoms at non-diagnostic levels. The Institutional Review Board at each contributing institution (i.e., Henry Ford Health Sciences Center, Howard University, Indiana University, SUNY Health Sciences Center at Brooklyn, University of California—San Diego, University of Connecticut Health Center, University of Iowa, and Washington University in St. Louis) reviewed and approved the protocols for genetic studies under which all participants were recruited. All participants gave written informed consent in accordance with the Declaration of Helsinki.

Participants completed a version of the Semi-Structured Assessment for the Genetics of Alcoholism (Bucholz et al., 1994), wherein lifetime DSM-IV substance dependence symptoms were assessed for alcohol, cannabis, cocaine, nicotine, and opioids. As genes influencing liability to substance use initiation may only partially overlap with genes influencing progression to various levels of dependence (Heath et al., 2002), categorical measures (five levels) for each substance were created to represent differential levels of involvement: (A) no lifetime (cannabis, cocaine, opioids) or non-regular (alcohol, nicotine) use, (B) non-problem use (i.e., use without endorsement of any dependence symptoms), (C) mild problems (i.e., 1–2 dependence symptoms), (D) moderate dependence (i.e., 3–5 dependence symptoms), and (E) severe dependence (i.e., 6–7 dependence symptoms; see Table 2 for distributions of participants across involvement levels). The lowest level of involvement was used as the reference group, though all groups were compared to one another (see Statistical Analyses). For cannabis, cocaine, and opioids, the reference group included individuals with no lifetime history of using the substance; for alcohol, those who had never drank at least once per month for 6 months or longer were considered to be minimally/not exposed, while for nicotine, this threshold was set at having smoked less than 100 cigarettes. The vast majority of individuals (82.9%) reported using multiple substances during their lifetime, with 17.0% reporting use of all substances assessed. Only 5.7% of the sample belonged to all substance-specific reference groups, reflecting no lifetime use of cannabis, cocaine, and opioids, and no regular use of alcohol and nicotine. Lifetime histories of problematic substance use also co-occurred, with 62.1% of the sample reporting at least one dependence symptom for two or more substances. Finally, in addition to alcohol dependence (46.9%), 17.4, 18.6, 50.9, and 6.9% of the sample endorsed 3 or more dependence criteria (unclustered) for cannabis, cocaine, nicotine, and opioid dependence, respectively (Table 2). A measure of general substance involvement (GENSUB) was generated by performing a confirmatory factor analysis (CFA) on the individual substance involvement measures in Mplus (v.7.11; Muthén and Muthén, 2015) and standardizing the resulting factor score (Supplementary Figure S2).

DNA was extracted from blood samples, and cell lines were developed as an additional DNA source. Samples were genotyped using lllumina Human1Mv1_CBeadChip at the Johns Hopkins Center for Inherited Disease Research (CIDR). Extensive and rigorous data cleaning was employed (Laurie et al., 2010), resulting in quality-controlled genotypic data for 948,658 SNPs (Bierut et al., 2010).

Polygenic risk scores were derived from the results of the PGC cross-disorder meta-analysis (CROSS; Cross-Disorder Group of the Psychiatric Genomics Consortium, 2013) of 5 psychiatric disorders (ADHD, AUT, BIP, MDD, SCZ). PRS were constructed for the following p-value thresholds based on the full GWAS summary statistics: 0.0001, 0.001, 0.01, 0.05, 0.1, 0.2, 0.3, 0.4, 0.5, and 1.0; these thresholds were selected to be consistent with the PRS analyses conducted in the PGC cross-disorder paper (Cross-Disorder Group of the Psychiatric Genomics Consortium, 2013). The PRS generation pipeline was coded in Python (v.2.7.6) using the Numerical Python (“NumPy,” v.1.7.1), StatsModels (v.0.5.0), and Python Data Analysis (“pandas,” v.0.12.0) libraries. Single Nucleotide Polymorphisms (SNPs) were required to have a MAF > 0.02, call rates > 0.98, and HWE p-values >10^-6 to be included in analyses. SNPs within the MHC region (chr6: 25000000:35000000) were excluded due to their complex patterns of linkage disequilibrium. All remaining SNPs were then pruned using p-value-informed clumping (i.e., grouping linked SNPs; R² = 0.10, 500 kb window), leaving 101,202 SNPs in SAGE for analysis. For each p-value threshold, the cross-disorder log odds-ratio for each component SNP was multiplied by the number of reference alleles for that SNP. These product terms were summed and divided by the total number of contributing SNPs, thus producing a single metric for each participant representing cross-disorder genetic vulnerability. These analyses were completed using the –score method in Plink (v.1.9; Chang et al., 2015). Individual disorder risk scores for the five psychiatric disorders (ADHD, AUT, BIP, MDD, SCZ) included in the cross-disorder meta-analysis were then generated in the same manner. Subsequent to the cross-disorder meta-analysis, a much larger second-generation GWAS of schizophrenia (SCZ2; Schizophrenia Working Group of the Psychiatric Genomics Consortium, 2014) was released, and analyses were therefore repeated for schizophrenia using SCZ2 PRS. Distributions of all PRS generated are presented in Supplementary Figure S1.

Associations between each thresholded CROSS PRS and GENSUB were tested using ordinary least squares regression. Multinomial logistic regression was then used to test associations across each level of involvement (i.e., no/non-regular use, non-problem use, mild problems, moderate dependence, and severe dependence) for specific substances and individual disorder PRS. Due to the large number of non-independent tests performed, an empirical significance threshold for α = 0.05 was determined using 10,000 label-swapping permutations (see Supplementary Materials and Methods for details). The lowest level of involvement was used as the reference group; thus, resulting odds-ratios (ORs) reflect increases or decreases in association for each level of substance involvement relative to the lowest level (i.e., no/non-regular use). Wald chi-square tests (for 1° of freedom) were used to examine whether the magnitude of these resulting ORs could be equated to each other and thus establish whether differences in PRS existed across involvement levels (e.g., comparison of the OR between no use and use with no problems vs. the OR between no use and use with 1–2 symptoms). To determine whether specific disorder-substance associations were driven by GENSUB, significant analyses were repeated with GENSUB as a covariate. Covariates across all analyses included sex, age quartiles, three ancestrally informative principal components, and study of origin.

The confirmatory one-factor model fit the data reasonably well in our sample (comparative fit index = 0.992; root mean square error of approximation = 0.106), supporting our proposed unidimensional conceptualization of alcohol, cannabis, cocaine, nicotine, and opioid involvement. Factor loadings were generally comparable across substances, though the loading for nicotine was somewhat lower (Supplementary Figure S2).

CROSS PRS were associated with increasing GENSUB factor scores (significant at 9 of 10 p-value thresholds; most significant at p < 0.5: β^∗ = 0.110, R² = 0.011, p < 0.001; Figure 1; Supplementary Table S1), indicating a positive relationship between genetic liability to multiple psychiatric disorders and general substance involvement.

Analyses of individual disorder PRS and specific substance involvement revealed several noteworthy associations (see Figure 2; Supplementary Tables S2–S7 for all associations), some of which remained significant when controlling for general substance involvement (i.e., GENSUB; Supplementary Tables S2–S7). We report all nominally significant associations below and note which survived permutation-based correction for multiple comparisons within the text and in Supplementary Tables S2–S7.

Associations between PRS and individual substances were only partially attributable to GENSUB, indicating specificity of certain relationships (e.g., ADHD PRS and nicotine involvement). This is significant, considering that twin studies implicate GENSUB as the primary source of genetic variance in individual substance use disorders (Kendler et al., 2003, 2007, 2012). Our factor loadings support this high degree of cohesiveness, with the possible exception of nicotine, also consistent with a prior twin study (Kendler et al., 2007). Notably, unlike prior research which has heavily relied on binary measures of substance use or dependence, we capitalized on the range of substance involvement present in our target sample due to the ascertainment strategy (Table 2), including non-problem use (i.e., use with no dependence symptoms) as well as multiple levels of problematic use (i.e., 1–2 vs. 3–5 vs. 6–7 dependence symptoms). Not only did this coding allow us to differentiate between early/casual and later/maladaptive levels of substance involvement, but it also allowed us to test whether PRS were associated with specific levels of substance involvement (i.e., to compare across groups). Despite factorial architecture suggesting only modest residual variance, several PRS remained associated with individual substances, particularly at non-problem or severe levels of involvement, even after accounting for this shared liability. This overall finding suggests that despite genetic susceptibility across substances being largely shared, variation at the extremes of the phenotype may be less well captured by measures such as GENSUB.

Some limitations of our study are noteworthy. First, comorbidities in the cross-disorder meta-analysis from which the PRS were derived, as well as in the target SAGE sample, may be subject to certain unmeasured confounds. For example, the PGC did not examine the extent of cocaine (or other substance) use in their sample population (Cross-Disorder Group of the Psychiatric Genomics Consortium, 2013), so our associations of SCZ PRS with severe cocaine dependence may be an artifact of increased cocaine use by people with schizophrenia (Serper et al., 1995), or of cocaine-induced psychosis resulting in a diagnosis of schizophrenia (Brady et al., 1991). Conversely, though severe psychopathology (i.e., AUT, BIP, SCZ) is likely to be uncommon in SAGE and unlikely to influence associations, more common psychopathologies (i.e., ADHD, MDD) were likely present. Therefore, associations with ADHD or MDD PRS may have been mediated by the actual expression of the disorder among SAGE participants (e.g., people with higher MDD PRS in SAGE develop MDD, which in turn is associated with substance use disorders). Though data on ADHD diagnosis in SAGE were unavailable, repeating MDD PRS analyses while controlling for DSM-IV MDD diagnosis did not alter results, indicating that associations between PRS and substance use outcomes were not directly related to disorder expression (Supplementary Table S8). Longitudinal studies of well-characterized populations, as well as an increased emphasis on the study of substance use disorders in consortia such as the PGC, will be critical to further address these questions of temporality and comorbidity.

Second, SAGE was ascertained for liability to substance dependence, specifically to alcohol, nicotine, and cocaine; the generalizability of these findings to the general population is thus unclear. Additionally, the factor structure of GENSUB might be somewhat sample-specific, and residual associations with involvement with specific substances (i.e., non-problem cannabis use) may have been artifacts of sample ascertainment. However, this ascertainment strategy allowed us to study the full range of substance involvement—from never-use to severe dependence—across both licit (i.e., alcohol, nicotine) and illicit (i.e., cannabis, cocaine, opioids) drugs, which would not have been possible in a population-based sample of comparable size. Nonetheless, it is important to replicate these findings in other samples.

Third, though multiple nominally significant relationships between genetic risk for individual psychiatric disorders and involvement with specific substances emerged, few survived correction for the large number of statistical tests performed. These results thus may represent spurious associations and should be interpreted with caution. However, given the consistency of certain associations (e.g., ADHD and nicotine use) with prior genetic (e.g., Chang et al., 2012) and epidemiological (e.g., Elkins et al., 2007; Lee et al., 2011) literature, they should not be dismissed outright. Notably, the only disorders with significant post-correction associations—MDD and SCZ—were those with the largest numbers of cases in the PGC cross-disorder meta-analysis (N_MDD = 9227, N_SCZ = 9379). Nominal associations may thus strengthen with larger discovery samples, which may provide more precise PRS estimates, as well as larger target samples. In support of this interpretation, repetition of schizophrenia PRS analyses with scores derived from the much larger second-generation PGC GWAS (N_SCZ2 = 36,989) yielded associations that were generally stronger than those from the first-generation analyses (Supplementary Tables S6 and S7; see Supplementary Figure S3 for comparison).

Fourth, while our PRS approach yielded evidence that shared common genetic architecture contributes to comorbidity between psychopathology and substance involvement, it does not provide insight into specific biological (e.g., reward-related neural responsiveness, epigenetically medicated changes in gene expression), psychological (e.g., anhedonia, impulsivity), and/or experiential (e.g., early life stress, peer group) mechanisms through which this risk is manifest (e.g., Olfson et al., 2014; Peña-Oliver et al., 2016; Ron and Barak, 2016). Compelling evidence suggests that psychopathology and substance involvement share overlapping neural systems (e.g., Buckholtz and Meyer-Lindenberg, 2012), molecular pathways (e.g., Ng et al., 2013), and environmental exposures (e.g., Kristjansson et al., 2016). Future PRS research probing biological systems and psychological traits common to both psychiatric and substance use disorders (e.g., Lancaster et al., 2016; Peña-Oliver et al., 2016) and incorporating environmental/experiential measures (e.g., French et al., 2015), alongside genomewide efforts to partition heritability into specific pathways and functional categories (e.g., Lee et al., 2012; Finucane et al., 2015), will further our understanding of the mechanisms underlying this comorbidity (Bogdan et al., 2016).