Beyond being useful for research purposes, researchers have begun to examine the potential of PRS to predict risk for medical outcomes in clinical settings. PRS for coronary artery disease , atrial fibrillation , type 2 diabetes , inflammatory bowel disease , and breast cancer have been found to be as predictive of these diseases as well known monogenic mutations, which tend to be rarer, and could lead to improved screening for larger numbers of individual who are at risk. Individuals in the top 5% of the PRS distributions had ~3 fold likelihood of having CAD, AF, T2D, IBS, or BC compared to the bottom 95%. For obesity, individuals in the top PRS decile were on average 13 kg heavier than those in the bottom decile. These studies demonstrate the potential for identifying individuals at heightened risk for various medical conditions using PRS. Given that AUD is a moderately heritable trait and GWAS for alcohol-related phenotypes are beginning to identify numerous variants associated with these outcomes, PRS for alcohol-related outcomes may be also able to identify individuals at heightened risk of developing an AUD. In the current analysis, we tested PRS in two target samples, a population-based sample and a clinically ascertained sample of families deeply affected by AUD, to evaluate the current state of alcohol-related PRS in relation to AUD and identifying those at heightened risk. We use several discovery samples from large-scale GWAS to create three PRS: a meta-analysis of two GWASs on alcohol-related problems, a recent large scale GWAS of alcohol consumption, and a GWAS for risky behaviors, including alcohol use. We chose to test PRS based on multiple alcohol-related GWAS because multiple lines of evidence indicate alcohol consumption and dependence have only partially shared genetic etiology. Additionally, we include a PRS for general risk behavior as there is robust evidence that the genetic risk for alcohol and other substance use disorders is shared with other disorders and behaviors related to reduced inhibitory control. Similar to recent work for specific medical conditions, greenhouse tables we compare the upper end of the PRS distribution at various thresholds to examine whether focusing on these upper parts of the distribution provide additional information in identifying those at increased risk of developing an AUD.
We acknowledge the exploratory nature of these analyses and the arbitrary nature of our thresholds in the absence of well-defined clinical risk scores, such as those for medical conditions like hypertension. Finally, we test the association of these PRSs with other substance use disorders , based on the robust finding that substance use disorders share an underlying genetic architecture, with the majority of the heritability shared across substances.We constructed lifetime criteria counts of cannabis, cocaine, and opioid use disorders based on DSM-5 criteria. We measured nicotine dependence criteria using the Fager strom Test for Nicotine Dependence , which assesses six criteria and has values ranging from 0 to 10 in both COGA and FT12. Because many illicit SUDs were not measured or rare in the FT12 data, we limit analyses of illicit SUD to COGA. Like AUD, these criteria counts represent the maximum reported for each respondent across the course of the study.In the case of FTND, this is limited to those who report smoking 100+ cigarettes in their lifetime.We created PRS derived from publicly available large scale GWASs. Information on genotyping and quality control is available in the Supplementary information. We created PRS using a Bayesian regression and continuous shrinkage method. PRS-CS uses LD information from an external reference panel to estimate the posterior effect sizes for each SNP in a given set of GWAS summary statistics. Both empirical tests and simulations have shown improved predictive power above traditional methods of score construction. For computational purposes, we limited the SNPS for score creation to HapMap3 SNPs that overlapped between the original GWAS summary statistics, the LD reference panel, and the target samples for score creation. We converted PRS to Z-scores for interpretation. We used four primary discovery GWASs to create three different PRSs. The first was from a recent GWAS of number of alcoholic drinks per week in approximately one million individuals provided by the GWAS & Sequencing Consortium of Alcohol and Nicotine Use. We obtained GSCAN summary statistics with all Finnish and 23 and Me cohorts removed . The PRS for alcohol problems were derived from a meta-analysis of two GWASs: a GWAS on the problem subscale from the Alcohol Use Disorders Identification Test in 121,604 individuals from the UK Biobank and the Psychiatric Genomcs Consurtium’s GWAS of alcohol dependence. Both FT12 and COGA were in the initial AD GWAS and we obtained summary statistics with each cohort removed . Finally, we derived a PRS for risky behaviors from a GWAS of the first prinicipal component of four risky behaviors from 315,894 individuals in the UK Biobank.
While this PRS does include alcohol consumption and smoking, it captures the shared variance between these substance use measures and the other two risky behaviors. These polygenic scores covered the domains of alcohol consumption , alcohol problems , and general externalizing .We first identified the predictive power for each PRS in both COGA and FT12 using the change in R2 above a baseline model with sex, age of last observation, the first ten ancestral principal components , genotyping array, and data collection site . We used linear/generalized-linear mixed-effects models with random intercepts to adjust for clustering at the family level and a pseudo-R2 for mixed models. In addition to the predictive power of individual PRS, we estimated the conditional effect of all PRS on AUD criteria to examine whether each PRS explained unique variance in AUD criteria. We also calculated the area under the curve of the conditional model containing all continuous PRS to estimate sensitivity/specificity. AUC provides an estimate of the probability a randomly selected case has predicted value more extreme than that of a randomly chosen control. An AUC of 0.5 indicates that a classifier does not provide any useful information in determining cases from controls . We next divided PRSs at several thresholds to examine whether there was a non-linear increase in risk of AUD across the PRS continuum. Finally, we compared mean values of other substance use outcomes for the top 5% in each PRS to those in the bottom 95%. We selected this threshold based in the increased prevalence of AUD in those in the top 5% of the PRS distributions . All code is available upon request from the corresponding author.In order to estimate whether individuals at the extreme end of the PRS distribution were at elevated risk of AUD, we compared the risk of AUD between those above and below a given threshold in the distribution. We divided these PRSs at the 80th, 90th, and 95th percentile in each sample and estimated the odds ratio for AUD in the top portion of the distribution relative to the bottom portion of the distribution . Table 2 provides the estimates for all of those models. Across each threshold for AUD severity in COGA, we observed a similar pattern where, as expected, those in the upper end of the polygenic distribution had greater odds of meeting criteria for AUD. However, regardless of the threshold, the OR’s at each threshold were roughly equivalent. For example, in the case of severe AUD, when dividing 80th percentile , 90th percentile , or 95th percentile , all of confidence intervals for the point estimates overlap. In FT12, there was a similar pattern. Though some of the point estimates appear to increase as the thresholds become more restrictive,vertical farming the confidence intervals again overlap.Researchers have begun to evaluate the potential for use of PRS in clinical settings. In this analysis, we examined the current predictive power and strength of association between several PRSs and a variety of SUDs, with a focus on AUD in both a clinically ascertained and a population-based sample. We were interested in which scores based on available GWASs provided the strongest association with alcohol use disorder, whether these scores explained unique variance in AUD in a conditional model, and how well these scores discriminated between cases and controls; what the risk of AUD was for those at the upper end of the risk continuum compared to the bottom; and 3) the levels of substance use disorder criteria for individuals at the top 5% of the polygenic score continuum compared to remaining 95%.In terms of which polygenic scores were the most predictive, we considered three scores: one based on problematic alcohol use , one based on alcohol consumption , and one based on general risky behaviors , as twin and family studies have shown alcohol and other risk behaviors to be genetically correlated traits. In both samples, the GSCAN DPW PRS was the most strongly associated, followed closely by the RISK PC PRS. When we included all of the PRS in one model, all three PRS were associated with AUD criteria in COGA. Only the RISK PC and GSCAN DPW PRS were associated with AUD criteria in FT12.
Overall, the unique contributions of each PRS reinforce the notion that the genetics of AUD are multifaceted, comprised of risk for level of consumption, alcohol-related problems, and behavioral disinhibition. Evaluating the AUC for the combined PRS revealed the combined effect of PRSs only marginally improved the AUC, similar to recent analyses for coronary artery disease and ischemic stroke. We ran a series of sensitivity analyses to test whether differences across the samples reflected age differences rather than differences in ascertainment. Restricting COGA to participants under 30 did not fundamentally change the results . Evaluating the AUC for the combined PRS revealed the combined effect of PRSs only marginally improved the AUC over models with just covariates. In an exploratory approach, we chose a series of more restrictive thresholds to divide the PRS distribution. The odds of having an AUD were statistically indistinguishable across each of the thresholds in both COGA and FT12. Even though the point estimates increased in some cases, the confidence intervals around these estimates were relatively large and they did not differ significantly. Additionally, there were only a small number of individuals in the severe category in FT12 and we urge caution in interpreting these estimates. Finally, the top 5% of the continuum for each PRS reported elevated rates of other SUD criteria compared to the bottom 95%. The RISK PC PRS was most associated with higher mean levels of SUD criteria, suggesting that risk for externalizing may be particularly useful in identifying individuals at risk for multiple SUDs. These initial findings suggest the current PRSs are unlikely to prove useful for SUDs in a clinical setting. Being able to eventually identify those at heightened risk for SUDs may allow for more targeted early intervention and prevention. However, before this is possible, larger discovery GWAS across substance use phenotypes with PRS that explain greater portions of the variance will be necessary. As GWAS sample sizes for SUDs increase, we will likely see increases in effect size. Additionally, using multivariate techniques to model the shared genetic architecture across existing SUD GWAS to include both aspects of externalizing and internalizing may also improve prediction. Inclusion of genetic data in a clinical setting will also require that psychiatrists and clinicians receive greater training in genetics and/or that they partner with genetic counselors, so they are both better able to understand what increased genetic risk means and be able convey that information accurately to their patients. In addition to clinical utility, we must ensure that regulations and protections surrounding the use of genetic information in clinical settings can adequately protect the rights of individuals who are identified to be “at risk.” This research has several important limitations. First, all analyses were limited to individuals of European ancestry because the discovery GWASs available were conducted in individuals of primarily European ancestry. It will be important to ascertain sizable samples of subjects with non European ancestries to properly estimate the predictive utility of PRS in non-European samples. This is especially important for racial-ethnic minorities so that health disparities are not further perpetuated. Second, our use of lifetime diagnoses may obscure the impact of changing genetic influences on the development of AUD across the life course.