Identifying risk factors involved in the common versus specific liabilities to substance use: A genetically informed approach

Abstract Individuals most often use several rather than one substance among alcohol, cigarettes or cannabis. This widespread co‐occurring use of multiple substances is thought to stem from a common liability that is partly genetic in origin. Genetic risk may indirectly contribute to a common liability to substance use through genetically influenced mental health vulnerabilities and individual traits. To test this possibility, we used polygenic scores indexing mental health and individual traits and examined their association with the common versus specific liabilities to substance use. We used data from the Avon Longitudinal Study of Parents and Children (N = 4218) and applied trait‐state‐occasion models to delineate the common and substance‐specific factors based on four classes of substances (alcohol, cigarettes, cannabis and other illicit substances) assessed over time (ages 17, 20 and 22). We generated 18 polygenic scores indexing genetically influenced mental health vulnerabilities and individual traits. In multivariable regression, we then tested the independent contribution of selected polygenic scores to the common and substance‐specific factors. Our results implicated several genetically influenced traits and vulnerabilities in the common liability to substance use, most notably risk taking (b standardised = 0.14; 95% confidence interval [CI] [0.10, 0.17]), followed by extraversion (b standardised = −0.10; 95% CI [−0.13, −0.06]), and schizophrenia risk (b standardised = 0.06; 95% CI [0.02, 0.09]). Educational attainment (EA) and body mass index (BMI) had opposite effects on substance‐specific liabilities such as cigarette use (b standardised‐EA = −0.15; 95% CI [−0.19, −0.12]; b standardised‐BMI = 0.05; 95% CI [0.02, 0.09]) and alcohol use (b standardised‐EA = 0.07; 95% CI [0.03, 0.11]; b standardised‐BMI = −0.06; 95% CI [−0.10, −0.02]). These findings point towards largely distinct sets of genetic influences on the common versus specific liabilities.


| INTRODUCTION
Substance use is a leading contributor to the global disease and disability burden 1 and is associated with high societal and economic costs. Of particular public health concern is the problematic use of multiple substances, such as the co-occurring use of cigarettes, alcohol and cannabis. This pattern of co-occurrence has pervasive longterm health implications. 2 During adolescence and emerging adulthood, the initiation of use of multiple classes of substances may be especially harmful, as it increases the risk of developing the clinical manifestation of a substance use disorder. 3 To inform prevention strategies, it is therefore essential to understand the origins of such problematic pattern of substance use.
According to the common liability model, the observed correlations between the use of different substances 2,4,5 can be explained by the presence of a common, nonspecific liability underlying the risk of use of different classes of substances. 6,7 Support for this model comes from several lines of research. For example, in observational studies, the use of different classes of substances is typically associated with a range of shared individual factors such as mental health vulnerabilities (e.g., schizophrenia, attention deficit and hyperactivity disorder [ADHD]), 8,9 personality traits (e.g., risk taking), 10,11 cognitive factors (e.g., educational attainment), 12 and physical characteristics (e.g., body mass index [BMI]). 13 Results from twin 4,14 and genomic studies 15,16 further indicate that the correlation between the use of different substances stems from a common liability that is largely genetic in nature.
Evidence regarding the common liability model from genomewide association studies (GWAS) is more challenging to interpret. So far, GWAS studies have most reliably identified single nucleotide polymorphisms (SNPs) that are associated with the use of particular classes of substances. 16,17 For example, a replicated finding is the association between the alcohol metabolism gene alcohol dehydrogenase 1B (ADH1B) and alcohol use 16,18 or the association between the nicotinic receptor gene CHRNA5 (cholinergic receptor nicotinic alpha 5 subunit) and cigarette use. 16 While this evidence appears to implicate only substance-specific genetic effects, recent powerful GWAS studies also identified SNPs with effects shared across two classes of substances (e.g., smoking and alcohol) and identified SNPs that extend beyond ADH1B and CHRNA5. 16 This highlights the importance of systematically modelling factors that reflect common versus substance-specific liabilities when assessing genetic influences on substance use.
Genome-wide findings also implicate that different substance use phenotypes share some polygenic liability with a number of individual traits and vulnerabilities, such as risk taking, 16,19,20 ADHD, 16,[20][21][22] depression, 21-23 neuroticism, 21 cognition 20,22 or schizophrenia. [20][21][22]24,25 This body of research suggests that the genetic architecture of the common liability may consist of highly polygenic and small indirect effects via a range of genetically influenced mental health vulnerabilities and individual traits. As such, if those traits and vulnerabilities are causally involved in the aetiology of the common liability to substance use, their respective genetic proxies (e.g., genetic variants associated with risk taking) must be associated with the common liability.
In this study, we propose to exploit the polygenic score (PGS) approach to further interrogate the aetiology of the common and substance-specific liabilities to substance use. A PGS is a continuous index of an individual's genetic risk for a particular phenotype, based on GWAS results for the corresponding phenotype. 26 PGSs can be used as genetic proxies indexing vulnerabilities and traits to study their role in the common and specific liabilities to substance use.
Employing PGSs as proxies for potential risk factors can be conceived as a first step in a series of genetically informed designs to strengthen causal evidence in observational studies. 27 For example, studies have used PGSs indexing a particular vulnerability or trait, such as depression or psychotic disorders, to test their association with the use of specific classes of substances including cannabis, 28 alcohol, 29,30 nicotine 29,30 or illicit substances. 29 However, this evidence does not provide insights regarding the aetiology of common versus substancespecific liabilities. One study has employed the PGS approach to study the effect of a few selected PGSs indexing mental health disorders on the use of multiple substances. 31 However, important traits and vulnerabilities previously implicated in the aetiology of substance use, including personality traits, cognitive measures and physical characteristics, remain to date untested.
We aimed to triangulate and extend previous phenotypic evidence by integrating genomic data with phenotypic modelling of the common versus specific liabilities to substance use in a longitudinal population-based cohort. We first generated 18 PGSs, indexing a range of genetically influenced mental health vulnerabilities and traits previously implicated in the aetiology of substance use. Second, we applied the PGS approach to test the association of the 18 genetically influenced vulnerabilities and traits with (a) a common liability to substance use capturing the co-occurrence of use of alcohol, cigarettes, cannabis and other illicit substances and (b) substance-specific liabilities that are independent of the common liability. By applying genetically informed methods such as the PGS approach to study refined phenotypes, this investigation has the potential to yield important insights for the aetiology of substance use and inform prevention and treatment programmes.

| Sample
We analysed data from the Avon Longitudinal Study of Parents and Children (ALSPAC). 32 Details about the study design, methods of data collection, and variables can be found on the study website (http:// www.bristol.ac.uk/alspac/). We used phenotypic data on substance use collected when the study participants were 17, 20 and 22 years of age. Genotype data were available for 7288 unrelated children of European ancestry after quality control (cf. Supporting information for details). Participants were included if they had at least one available substance use measure across the three time points, resulting in a final sample of 4218 individuals.

| Summary statistics datasets
We collected summary statistics from 32 publicly available GWAS derived from discovery cohorts, which did not include ALSPAC participants (Table S2), indexing domains such as mental health vulnerabilities (e.g., depression), personality (e.g., risk taking), cognition (e.g., educational attainment), physical measures (e.g., BMI) and substance use (i.e., nicotine, alcohol and cannabis use). We chose GWAS indexing either substance use behaviours or individual traits and vulnerabilities that could be plausibly linked to substance use (cf. Section 1). From the initial 32 GWAS, we only included those with a sufficiently large sample (N > 20 000 participants) and we excluded several GWAS to avoid content overlap, resulting in a final selection of 18 GWAS summary statistics (cf.

| PGS analysis
Eighteen PGSs were generated utilising PRSice software version 2.2 (http://www.prsice.info/), 26 based on ALSPAC genotype data and the selected GWAS summary statistics. The PGSs for each individual were calculated as the sum of alleles associated with the phenotype of interest (e.g., schizophrenia), weighted by their effect sizes found in the corresponding GWAS. Clumping was performed in order to remove SNPs in linkage disequilibrium (r 2 > 0.10 within a 250-bp window). The PGSs were generated using a single p-value threshold of 1 in order to limit multiple testing while maximising the potential predictive ability of the PGSs. 36 2.3.2 | Trait-state-occasion models of substance use All analyses were conducted in R version 3.5.1 using the 'Lavaan' package. 37 First, trait-state-occasion (TSO) structural equation models were fitted using the scores for cigarette, alcohol, cannabis and other illicit substance use at each time point. 38 This approach enabled us to model latent factors of substance use that are stable over time, including (a) a common factor of all substances and (b) substance-specific factors. Such advanced phenotypic modelling retains a higher degree of precision and specificity compared with simple observed substance use phenotypes. Missing data on the substance use indicators were handled using full maximum likelihood estimation. The model parameters were estimated using robust standard errors due to nonnormality of the substance use scores. The TSO model was tested using available model specifications. 39 Further details are provided in the Supporting information and in Figure 1. Second, we tested the associations of each PGS with both the common and substance-specific latent factors (single-PGS TSO models) in order to explore their individual effects. False discovery rate (FDR) corrected p values 40 are provided to account for multiple testing. Finally, we tested two sets of multivariable TSO models (multi-PGSs TSO models) for each latent factor, in which we included only those PGSs that remained significant after FDR correction. In the first set, we included PGSs indexing substance use phenotypes (i.e., PGSs indexing dependency and frequency of cigarette, cannabis and alcohol use). In the second set, we included PGSs indexing mental health vulnerabilities and traits. The aim of this multivariable approach was to assess the independent effect of each PGS, controlling for potential pleiotropic effects (i.e., association of a single PGS with an outcome explained by its genetic overlap with other PGSs). All PGS-regression models were included directly within the TSO models. An example of the Lavaan syntax used for the single and multi-PGSs models can be found in the Supporting information.
All regression models were controlled for sex and population stratification by including 10 principal components as covariates. All PGSs were standardised.

| RESULTS
The descriptive statistics of substance use in our sample can be found in Table S4 Table S6.

| Effects of the PGSs reflecting substance use
The standardised regression coefficients and confidence intervals of the associations of the PGSs with the common and substance-specific factors are shown in Figure 3 (cf. Tables S7 and S8). As expected, the factors capturing cigarette and alcohol use were predicted by their respective PGSs (e.g., frequency of cigarette/alcohol use), reflecting specific genetic effects (e.g., linked to substance-specific metabolism).
The common factor was independently predicted by two substance use PGSs (age of onset of cigarette use and alcohol frequency), in line with evidence implicating age of onset of cigarette use as a liability marker for initiation of use of other substances. 41 Other substancespecific factors were not predicted by their respective PGSs

| Substance-specific factor: Cigarette use
In the single-PGS TSO models, five PGSs were associated with the cigarette use factor following FDR correction (educational attainment,

| Substance-specific factor: Cannabis use
None of the PGSs was associated with the cannabis use factor.

| Substance-specific factor: Other illicit substance use
In the single-PGS TSO models, five PGSs were associated with the factor representing other illicit substance use following

| DISCUSSION
This study is the first genomic investigation using the PGS approach to examine the contribution of a range of individual traits and vulnerabilities to both common and specific liabilities to substance use. We

| Insights for the aetiology of substance use
In this study, we exploited the PGS approach as a genetically informed method 43 to strengthen inference on risk and protective factors involved in liabilities to substance use, thereby enabling triangulation of previous phenotypic evidence with distinct sources of bias (e.g., traditional observational evidence). Using the PGS approach, our results helped to tease apart some of the genetic predispositions (e.g., PGS indexing schizophrenia liability) that indirectly contribute to common and substance-specific liabilities to substance use. In particular, different sets of genetically influenced mental health vulnerabilities and traits are likely to be involved in common versus substancespecific liabilities. Importantly, all associations found in this study can be conceptualised as indirect effects of genetically influenced traits and vulnerabilities. To illustrate, our findings suggest that a genetic liability to risk taking could lead to greater risk-taking behaviour, which in turn could affect an individual's propensity to engage in substance use irrespective of the class of the substance. However, it should be noted that the PGS approach relies on a number of key assumptions (see Section 4.5). As such, we cannot rule out the possibility that confounders impact on the associations between PGSs and our substance use outcomes.

| Risk and protective factors involved in the common liability to substance use
Our results confirm previous findings of a common liability that partly underlies the use of different classes of addictive substances, such as cigarettes, alcohol, cannabis and other illicit substances. 6,44 Regarding its origins, our findings reveal that a genetic liability to high risk taking, low extraversion and schizophrenia contributes to the common liability to substance use. This corroborates previous phenotypic evidence, which reported associations between substance use and similar traits and vulnerabilities. 8,10,11,45 Intriguingly, a genetic predisposition for risk taking was most robustly associated with a common liability to substance use, but only to a lesser extent with substance-specific liabilities (cf. next paragraph). This indicates that individuals susceptible to risk taking are more likely to use an array of different substances, irrespective of their class. Similarly, a genetic predisposition to extraversion was most strongly associated with the common liability to substance use, whereas its associations with substance-specific liabilities were weaker. Thus, high extraversion may protect against the use of various substances. Furthermore, the common liability was influenced by genetic risk for schizophrenia. Taken together, these findings are in line with the notion that the use of various substances could partly reflect a self-medication strategy for those individuals more vulnerable to psychopathology and maladaptive personality traits. 46 This is in line with theories implicating the reward system as a common pathway underlying the use of multiple substances-a system altered in distressed individuals and for whom the use of substances may represent a mean to restore homeostasis. 47 Finally, our results suggest that shared genetic effects among different substances of use are substantially polygenic in nature, involving many genetic variants exerting indirect and small effects (e.g., polygenic association via risk taking). Future large GWAS may therefore benefit from modelling a common liability to substance use, similar to recent genome-wide attempts aiming to identify common genetic variation underlying psychiatric traits. 48,49

| Risk and protective factors involved in substance-specific liabilities
Our results also showed that a substantial proportion of the phenotypic variation in substance use could not be explained by a common liability. Using the PGS approach to identify genetically influenced risk and protective factors involved in the substance-specific liabilities revealed three patterns of associations. First, (a) we identified a set of factors that were linked to both the common liability to substance use, as well as to substance-specific liabilities. Second, (b) several factors were linked to substance-specific liabilities but did not contribute to the common liability. Third, (c) some traits previously implicated in substance use were not associated with any of the substance-specific liabilities.
Regarding (a), we found that all factors involved in the common liability including a genetic predisposition for risk taking, extraversion and schizophrenia also contributed to the liability to alcohol use.
Hence, the aetiologies of these two liabilities (i.e., alcohol vs. common) are partly based on overlapping risk factors. At the same time (b), our results showed that two individual traits-BMI and educational attainment-were not linked to the common liability but predicted substance-specific liabilities. Interesting results emerged regarding the direction of the identified associations. For example, we found that a predisposition for high educational attainment increased the risk of alcohol and illicit substance use but reduced the risk of cigarette use. This is consistent with the notion that education makes people less likely to smoke cigarettes 50 due to an increased knowledge of its adverse health consequences. At the same time, greater education may provide more opportunities to consume alcohol and access other substances, as indicated by previous observational evidence. 51 Opposite effects were also present for BMI. Here, a genetic predisposition for high BMI increased the risk of cigarette use, while reducing the risk of alcohol and other illicit substance use. The same pattern of associations has been reported in observational studies.
For example, compared with normal weight adolescents, obese adolescents were at reduced risk of alcohol and illicit substance use, but had an elevated risk of cigarette use. 13 As nicotine is known to suppress appetite, this may suggest that adolescents with a greater predisposition to high BMI could smoke more in an attempt to control their appetite. 52 Finally (c), some of the previously implicated risk factors (e.g., neuroticism and ADHD) 9,10 were not associated with the common or substance-specific liabilities in our sample. First, this could reflect a lack of power of the PGSs used in the analysis. However, we used powerful PGSs (e.g., neuroticism, derived from a GWAS with N > 160 000) that have been shown to predict rare outcomes in comparable samples. 53 Second, some PGSs were associated with substance use liabilities only in less controlled models (e.g., ADHD and depression predicting other illicit substance use only in single-PGS but not multi-PGSs models). In addition to power issues, this may indicate that the effects of ADHD/depression were explained by potentially co-occurring traits that we included in our multivariable models. lowing interventions targeting abilities related to risk taking (e.g., selfregulation) in adolescents. 54 Our results also highlight that it is important to target those individuals at greatest risk of developing a problematic pattern of substance use based on pre-existing vulnerabilities such as schizophrenia. Hence, in adolescents with prodromal symptoms, particular emphasis may need to be placed on the prevention of substance use. Finally, it is important to better understand the mechanisms underlying some of the substance-specific associations found in this study (e.g., high BMI as a risk factor for cigarette use) in order to design more effective prevention and intervention strategies.

| Limitations
By using genetic proxies that are more robust to confounding, 27

| CONCLUSION
Our findings reveal that distinct sets of genetically influenced vulnerabilities and protective factors are likely to be involved in the common versus substance-specific liabilities to substance use. In particular, a genetic predisposition to high risk taking, low extraversion and schizophrenia may be associated with the individual's susceptibility to the use of any type of substance. Additionally, genetic predispositions related to educational attainment and BMI were related to the use of multiple specific substances, although in opposite directions. Prevention programmes in adolescents may benefit from focusing on these vulnerabilities and protective factors.

AUTHOR CONTRIBUTIONS
Iob, Schoeler and Pingault had full access to all of the data in the study and take responsibility for the integrity of the data and the accuracy of the statistical analyses.

ROLE OF THE FUNDER/SPONSOR
The funding sources had no role in the design and conduct of the study; collection, management, analysis and interpretation of the data; preparation, review or approval of the manuscript; and decision to submit the manuscript for publication.

ADDITIONAL CONTRIBUTIONS
We are grateful to all the families who took part in this study, the midwives for their help in recruiting them, and the whole ALSPAC team, which includes interviewers, computer and laboratory technicians, clerical workers, research scientists, volunteers, managers, receptionists and nurses.