A Developmentally-Informative Genome-wide Association Study of Alcohol Use Frequency

Abstract

Contemporary genome-wide association study (GWAS) methods typically do not account for variability in genetic effects throughout development. We applied genomic structural equation modeling to combine developmentally-informative phenotype data and GWAS to create polygenic scores (PGS) for alcohol use frequency that are specific to developmental stage. Longitudinal cohort studies targeted for gene-identification analyses include the Collaborative Study on the Genetics of Alcoholism (adolescence n=1,118, early adulthood n=2,762, adulthood n=5,255), the National Longitudinal Study of Adolescent to Adult Health (adolescence n=3,089, early adulthood n=3,993, adulthood n=5,149), and the Avon Longitudinal Study of Parents and Children (ALSPAC; adolescence n=5,382, early adulthood n=3,613). PGS validation analyses were conducted in the COGA sample using an alternate version of the discovery analysis with COGA removed. Results suggest that genetic liability for alcohol use frequency in adolescence may be distinct from genetic liability for alcohol use frequency later in developmental periods. The age-specific PGS predicts an increase of 4 drinking days per year per PGS standard deviation when modeled separately from the common factor PGS in adulthood. The current work was underpowered at all steps of the analysis plan. Though small sample sizes and low statistical power limit the substantive conclusions that can be drawn regarding these research questions, this work provides a foundation for future genetic studies of developmental variability in the genetic underpinnings of alcohol use behaviors and genetically-informed, age-matched phenotype prediction.

Introduction

Frequent heavy alcohol use can lead to alcohol use disorder (AUD), which accounts for three million deaths and over 133 million life years lost to disability and death worldwide per year (WHO, 2018). Modest levels of alcohol consumption are also associated with health consequences, especially in younger age demographics (Bryazka et al. 2022). Alcohol use is under significant genetic influence, with genetic factors accounting for approximately 50% of the variability in risk for heavier alcohol consumption (Verhulst et al. 2015). Accordingly, many studies have been conducted to identify the underlying genetic variants that contribute to the propensity for alcohol use.

Genome-wide association studies (GWAS) adopt an agnostic approach to examine the entire genome for association with a phenotype in large samples (Visscher et al. 2017). Large sample sizes are needed for GWAS because of the polygenicity of complex traits and the low effect size of each individual single nucleotide polymorphism (SNP). Few studies have the data required to model developmental considerations in GWAS, instead relying on phenotype definitions that pool across time and developmental periods to maximize the number of samples that can be included in a single meta-analysis. This approach may not accurately model the ways that genetic influences on complex psychiatric and behavioral phenotypes vary across development.

A robust body of research using genetically-informative twin data demonstrates that genetic influences on alcohol use vary across development. Heritability changes across development; for instance, previous studies have found that the heritability of alcohol use increases from adolescence to emerging adulthood (Huibregtse et al. 2016; Kendler et al. 2008; Rose et al. 2001; Viken et al. 1999). This increase in heritability over time is partly driven by “genetic innovation”, which refers to new genetic risk factors that emerge throughout development (Edwards & Kendler, 2013; Long et al. 2017). Simultaneously, other genetic risk factors may become less important at later ages. Different genetic variants influence alcohol use at different developmental periods. Twin studies indicate that alcohol-specific and broader externalizing genetic factors operate at different developmental stages (Kendler et al. 2011; Meyers et al. 2014). The environment plays an important role in shaping these differences throughout development; for example, alcohol use in adolescence requires illicit access to alcohol. Relatedly, broad externalizing genetic factors are particularly salient in adolescent alcohol use (Kendler et al. 2011; Meyers et al. 2014).

The omission of developmental considerations from GWAS has downstream consequences for statistical analyses that leverage GWAS summary statistics for genetic prediction of phenotypes with polygenic scores (PGS), reducing the effectiveness of PGS in younger target samples. For example, PGS derived from developmentally-agnostic GWAS (Liu et al. 2019) predict 0.58% and 0.61% of the variance in alcohol consumption in adolescence and early adulthood, respectively (Kandaswamy et al. 2021), but predict 2.4% of the variance in alcohol consumption in an older target sample (age 24-32; Liu et al. 2019). A recent study that examined age-specific effects of an alcohol consumption PGS from an adult discovery sample (Kranzler et al. 2019) found that the PGS were associated with alcohol use in adulthood, but not adolescence (Elam et al. 2021). Similar patterns may be expected for PGS that are derived from adolescent samples and applied to adult samples. Despite previous findings that the nature and magnitude of genetic effects vary throughout development (Aliev et al. 2015; Dick et al. 2006; Edwards & Kendler, 2013; Kendler et al. 2011; Meyers et al. 2014; Sakai et al. 2010), PGS have not been constructed to model this variability.

The first aim of this project was to advance gene discovery by building models for gene identification that incorporate developmental changes in genetic influences on alcohol use across three developmental periods (adolescence, age 12-17; early adulthood, age 18-25; adulthood age 26+). To achieve this, we began by conducting a common factor GWAS of alcohol use frequency at different developmental stages using a meta-analytic approach in Genomic SEM (Grotzinger et al. 2019). The hypothesis for this analysis was that the meta-analyzed GWAS results would yield unique associations at different developmental stages. A diagram of the genomic structural equation model is depicted in Figure 1. To support the interpretation of this analysis, genetic correlations were calculated between the components of the genomic structural equation model and a series of other phenotypes using publicly available GWAS summary statistics.

Figure 1.

Common Factor Genomic Structural Equation Model

Genetic variance shared between all three developmental periods is indexed in the common factor. Genetic variance that is distinct from the common factor is indexed in the residuals. Model identification was achieved by fixing the variance of the common factor to one. Factor loadings are labeled with lambda. Residual variances are labeled with epsilon. The model is saturated.

The second aim of this project was to leverage results from the longitudinal GWAS for genetic prediction of age-matched alcohol use outcomes in an independent sample. We constructed PGS using weights from the residual components (U_12-17, U_18-25, U₂₆₊) and the common factor (η_common) of the genomic structural equation model. The hypothesis for this analysis was that the age-specific PGS would predict their corresponding alcohol use phenotype significantly better than a common factor PGS from the same model.

This project represents a preliminary study of how developmental considerations can be incorporated into genome-wide association studies, drawing on previous applications of Genomic SEM to model change and stability in genetic effects across age groups (Gillespie et al. 2022) and the residual genetic variance of a phenotype after accounting for a common factor (Barr et al. 2022). Gene discovery analyses require large sample sizes and the statistical power of the current work is limited. Thus, the third aim of this project was to estimate the required sample size for an adequately powered implementation of the model. We conducted a power analysis using simulation to estimate the sample size required to detect SNP effects on residual genetic variance in adolescence. The strategy of disaggregating developmental periods in genome-wide association studies is promising and can reveal new relationships, although it will require large sample sizes.

Methods

Samples

This project used three longitudinal cohort studies for developmentally-informed gene discovery: The Collaborative Study on the Genetics of Alcoholism (COGA), The Avon Longitudinal Study of Parents and Children (ALSPAC), and The National Longitudinal Study of Adolescent to Adult Health (Add Health). In order to model developmentally-salient genetic effects, data were aggregated within developmental periods by taking the maximum value of available data for each participant within the following age ranges: ages 12-17 (adolescence), ages 18-25 (early adulthood), and ages 26+ (adulthood). In alignment with previous work suggesting that genetic liabilities for initiation and use of alcohol are distinct (Heath et al. 1991a; Pagan et al. 2006; Fowler et al. 2007), all analyses excluded lifetime non-drinkers. Additionally, all analyses were limited to European ancestry participants to limit the confounding influence of population stratification.

COGA is a multi-generational family-based study of genetic and environmental factors for alcohol use disorder and related traits, which ascertained alcohol-dependent probands and a smaller number of comparison families from six US sites (Begleiter, 1995; Dick et al. in press). The current work focused on the initial COGA sample of proband and control families (Phase 1 and Phase 2), as well as the COGA Prospective Study (Phase 4). For Phase 1 and Phase 2, probands from alcohol treatment centers and their families were invited to participate if the family had two or more members in the COGA catchment area. Comparison families were recruited from the same communities. Phase 4 examines how genetic and environmental risk unfolds across development among offspring of the initial COGA sample. Offspring between ages 12-22 with at least one parent who had previously completed an interview were assessed every two years (Bucholz et al. 2017; Dick et al. in press). The Institutional Review Boards at all data gathering sites approved this study, and written consent (and assent for adolescents) was obtained from all participants. Reports between ages 12-17 (n=1,118, 49.73% female), 18-25 (n=2762, 52.28% female), and after age 26 (n=5,255, 53.80% female) were collected in Phase 4, Phase 1, and Phase 2 assessments as part of the SSAGA interviews (Bucholz et al. 1994).

Add Health is a nationally representative longitudinal cohort of adolescents who were in grades 7-12 in the US in 1994-95 (Harris, 2013). Participants were identified from 132 schools. The cohort has been followed through the transition to adulthood in five waves, with data collection occurring via in-home interviews. The sample includes siblings. Measures are primarily focused on causes of adolescent health behavior in the multiple contexts of adolescent life. Genetic data were collected in Wave 4. Add Health participants provided written informed consent to participate in the study in alignment with University of North Carolina School of Public Health Institutional Review Board guidelines. Reports between ages 12-17 (n=3,089, 53.71% female) were collected in Wave 1 and Wave 2. Reports between ages 18-25 (n=3,993, 54.30% female) were collected in Wave 3 and Wave 4. Reports after age 26 (n=5,140, 52.98% female) were collected in Wave 3, Wave 4, and Wave 5.

ALSPAC is a large longitudinal birth cohort which includes reports from approximately 14,000 children and their parents from early in the mothers’ pregnancy through childhood, adolescence, and emerging adulthood (Boyd et al. 2013; Fraser et al. 2013; Northstone et al. 2019). Pregnant women residing in Avon, UK with expected dates of delivery between 1st April 1991 and 31st December 1992 were invited to take part in the study. The total sample size for analyses using any data collected after the age of seven is therefore 15,447 pregnancies, resulting in 15,658 fetuses. Of these, 14,901 children were alive at 1 year of age. The project has collected comprehensive health-related information, including phenotypic outcomes, environmental factors, and DNA, with >85 assessments from mothers, their partners, and children, conducted from the prenatal stage through emerging adulthood at yearly, or more frequent, intervals. Please note that the study website contains details of all the data that is available through a fully searchable data dictionary and variable search tool (http://www.bristol.ac.uk/alspac/researchers/our-data/). Study data were collected and managed using REDCap electronic data capture tools hosted at the University of Bristol (Harris et al. 2009). REDCap (Research Electronic Data Capture) is a secure, web-based software platform designed to support data capture for research studies. Ethical approval for the study was obtained from the ALSPAC Ethics and Law Committee and the Local Research Ethics Committees (http://www.bristol.ac.uk/alspac/researchers/research-ethics/). Informed consent for the use of data collected via questionnaires and clinics was obtained from participants following the recommendations of the ALSPAC Ethics and Law Committee at the time. The current work focuses on offspring assessments in adolescence and early adulthood. Reports between ages 12-17 (n=5,382, 53.57% female) were collected at ages 12.5, 13.5, 15.5, 16, and 17.5 from offspring. Reports between ages 18-25 (n=3613, 60.75% female) were collected at ages 18 and 20 from offspring.

Measures

Alcohol Use Frequency

Measures of alcohol use frequency varied between and within the 3 samples of interest. Ordinal survey items measuring the frequency of alcohol use were transformed to a pseudo-continuous scale of drinking days per year. Ordinal frequency categories were converted to the median of the described range and rescaled from the original scale of measurement to reflect drinking days per year by multiplying by the corresponding constant (i.e., days per week * 52; days per month * 12). A similar procedure has been used in previous studies to construct pseudo-continuous measures of alcohol consumption (Dawson, 2000; Salvatore et al. 2016; Thomas et al. 2018). The resulting transformed variables have a fixed range of 0 to 365, precluding the inclusion of extreme outlier values. The maximum value was taken for participants who had multiple observations in a single developmental period. A complete description of survey items that were aggregated under each developmental period can be found in Table S1.

Genotyping

In COGA, participants were genotyped on four different genotyping arrays: the Illumina 1M, Illumina OmniExpress 12VI, and Illumina 2.5M (Illumina, San Diego, CA), and Smokescreen (BioRealm LLC, Walnut, CA); details of genotyping, imputation and quality control procedures are available in Lai et al. (2019) and Johnson et al. (in press). In ALSPAC, participants were genotyped on the Illumina Human Hap550 Quad array. Additional information regarding genotyping procedures in ALSPAC is available in Paternoster et al. (2011) and Chong et al. (2021). In Add Health, participants were genotyped on the Illumina Human Omni1-Quad BeadChip and the Illumina Human Omni-2.5 Quad BeadChip. Genotypes in Add Health were imputed together using a set of 609,130 SNPs in common between the two arrays. Additional information regarding genotyping procedures in Add Health is available in Highland et al. (2018).

Covariates

Covariates included in all analyses were sex, and 10 ancestry principal components. Sex was assessed using a single binary item in each study. Analyses in the COGA and Add Health samples included birth year as a covariate to control for generational birth cohort differences. All participants from ALSPAC were born in 1991 or 1992. Age at assessment was calculated from date of birth (Table S2). Covariates in COGA also included genotyping array, and in early adulthood and adulthood, individual proband status. We did not include age as a covariate because the aim of this study was to examine genetic effects in different age groups and because of the of the potential for collinearity between age at assessment and birth year.

Analysis Plan

A schematic of the full analysis plan is included in Figure 2 and described below.

Figure 2.

Diagram of the analysis plan in ALSPAC, COGA, and Add Health

The steps of the analysis plan include data preprocessing, within-sample GWAS, GWAS meta-analysis, estimation of SNP-based heritability and genetic covariances by LD-Score Regression, genomic structural equation modeling with SNP effects, calculation of genetic correlations with other phenotypes, and polygenic risk scoring.

Developmentally-Informative GWAS

Standard pre-imputation quality control metrics (Marees et al. 2018) were applied in all within-sample genetic analyses including removing SNPs with minor allele frequency (MAF) <1%, call rate <95%, Hardy-Weinberg Equilibrium (HWE) p < 1e-6, and removal of participants with excess autosomal heterozygosity or homozygosity (F < −.1 or F > .1). All samples were imputed to the Haplotype Reference Consortium (HRC) reference panel (McCarthy et al. 2016). SNPs with low imputation quality (Info Score < .80) were removed prior to analysis.

Separate GWAS were conducted at each developmental period for each available sample with linear regression using PLINK (Purcell et al. 2007) in unrelated participants (ALSPAC). We conducted GWAS with a linear mixed model with a genomic relatedness matrix (GRM) using Genome-wide Complex Trait Analysis (GCTA; Yang et al. 2011) in COGA and Add Health to adjust the standard error of SNP effect sizes for non-independence of observations among participants that are biological relatives. GRMs were constructed in each sample (COGA, Add Health) using GCTA. SNPs were pruned for GRM calculation using a 50 SNP window, shifting by 5 SNPs, with a variance inflation factor threshold of 2 in PLINK. GWAS summary statistics were meta-analyzed within developmental periods using an inverse variance-weighted meta-analysis in METAL (Willer et al. 2010).

Linkage disequilibrium score regression (LDSC; Bulik-Sullivan, Loh, et al. 2015), was performed on the meta-analyzed GWAS summary statistics to construct a genetic covariance matrix using the Genomic SEM package (Grotzinger et al. 2019). Previous work indicates that LDSC is robust to sample overlap (Bulik-Sullivan, Finucane, et al. 2015), facilitating the analysis of repeated measures across developmental periods. LD scores were drawn from the 1000 Genomes reference panel (McVean et al. 2012) and restricted to HapMap3 SNPS (Altshuler et al. 2010), which are well characterized in terms of LD structure. 1,193,613, 1,193,617, and 1,170,827 HapMap3 SNPs were included in the LD score regression after matching meta-analyzed summary statistics to the reference panel and LD scores for adolescence, early adulthood, and adulthood, respectively. We fit a common factor model to the genetic covariance matrix using the Genomic SEM package (Grotzinger et al. 2019) in R (R Core Team, 2017). A diagram of the common factor genomic structural equation model (gSEM) is included in Figure 1. The variance of the common factor was fixed to 1 to identify the model without SNPs. Loadings from the common factor to genetic variance at each developmental period (λ_12-17, λ_18-25, λ₂₆₊) and the corresponding residual variances (U_12-17,U_18-25,U₂₆₊) were freely estimated.

Next, we estimated SNP effects on the components of the gSEM. The tolerance setting for matrix inversion was set to a relatively liberal value (1e-50) to allow model fitting to proceed. The ‘standard’ option was selected to implement Genomic Control. SNPs were restricted to those appearing in both the meta-analyses and the 1000 Genomes reference panel and had a MAF > .01 in the reference panel. The total number of SNPs included in the analysis was 6,707,536. The average of SNP-specific sample sizes for adolescence, early adulthood, and adulthood were 8,869.29, 9,647.64, and 9,894.18, respectively. The effective sample size for the common factor, calculated using the formula defined in Mallard et al. (2022a), was 12,180.07.

We considered two parameterizations of the gSEM GWAS model. The gSEM GWAS model for residual genetic variance in each developmental period was fit with paths from the SNP to both the residual and the common factor. This approach partitions the SNP effects into a component that is attributable to the common factor and a component that is directly associated with the residual genetic variance. An alternate approach models total genetic variance in each developmental period by omitting the path from the SNP to the common factor. A simulation supporting this distinction between the two model parameterizations is provided in the supplemental material. A diagram of the gSEM GWAS model for SNP effects on the total genetic variance in adolescence is included in Figure S1. A diagram of the gSEM GWAS model for SNP effects on the common factor is included in Figure S2.

Statistical power was not adequate to estimate genetic correlations between the model residuals and a series of phenotypes from other studies. Interpretable heritability estimates have a range of zero (0% heritable) to one (100% heritable); however, LDSC heritability estimates are not bound to the range of interpretable values. LDSC heritability estimates can sometimes fall below this range and be negative if the true value of the heritability is close to zero or statistical power is low enough that sampling variance produces a point estimate below zero. Negative heritability estimates are not interpretable and prevent the calculation of genetic correlations with the phenotype that has negative heritability. The residual parameterization produced negative heritability estimates in early adulthood and adulthood (early adult H²_SNP = −6e-04, adult H²_SNP = −4e-04), preventing the calculation of genetic correlations. The total variance parameterization of the gSEM GWAS model, which omits the path from SNP to the common factor, was used for this component of the analysis plan. Negative LDSC heritability estimates do not prevent polygenic risk scoring. The residual parameterization was used to construct polygenic scores that are specific to developmental periods. These two approaches to modeling are not directly comparable and the results for the discovery analysis and prediction by polygenic scores are considered separately.

For the GWAS in each developmental period (U_12-17, U_18-25, U₂₆₊), paths were estimated from each SNP to the corresponding model component. Genetic correlations were calculated between the components of the genomic structural equation model and a set of other phenotypes with publicly available GWAS summary statistics (Neale Lab 2018; Walters et al. 2018; Karlsson Linnér et al. 2019, 2020; Sanchez-Roige et al. 2019; Howard et al. 2019; Purves et al. 2020; Okbay et al. 2022; Saunders et al. 2022) using LDSC. A list of the phenotypes with sample sizes and citations is included in Table S3.

Validation Analysis: Polygenic Scores

An alternate version of the Aim 1 discovery analysis was conducted with COGA removed in order to facilitate PGS construction in the COGA sample. COGA was selected for PGS validation to maximize the remaining sample size in the meta-analytic GWAS. The residual variance parameterization, which includes the path from SNP to the common factor, was used for this component of the analysis plan to construct polygenic scores that are specific to developmental stages. The total number of SNPs in the analysis, included in the meta-analysis of ALSPAC and Add Health and the reference panel, was 6,071,632. Sample sizes for the gSEM GWAS analysis in adolescence, early adulthood, adulthood, and the common factor with COGA removed were 7,869.59, 7,070.74, 5,149, and 8230.73, respectively. A diagram of the gSEM GWAS model for residual genetic variance in adolescence is included in Figure S3.

PGS were constructed from the gSEM GWAS of the residuals and common factor using PRS-CS (Ge et al. 2019). The ϕ global shrinkage parameter was set to 1e-2 to accommodate the small sample size of the meta-analyzed GWAS in adolescence, early adulthood, and adulthood. All PGS were scaled to a standard normal distribution with a mean of 0 and a standard deviation of 1. A series of mixed effects models with random intercepts for family group were used to assess the effect of the age-specific PGS and the common factor PGS on the corresponding phenotype in COGA using the lme4 (Bates et al. 2015) and lmerTest packages (Kuznetsova et al. 2017) in R (R Core Team, 2017). A separate mixed effects model was used for each developmental period included in the gSEM GWAS. Sex, birth year, genotyping array, and 10 PCs were included as covariates in each model. Individual proband status was included as a covariate in early adulthood and adulthood. The Benjamini-Hochberg procedure for controlling the False Discovery Rate (FDR) (Benjamini & Hochberg, 1995) was implemented to correct p-values for three tests using the p.adjust function in R (R Core Team, 2017) with option “BH”. FDR corrected confidence intervals were constructed by recalculating standard errors as a function of the FDR corrected p-value.

Pseudo-R² (Nakagawa, & Schielzeth, 2013) and model fit indices were obtained using the performance package (Lüdecke et al. 2021) in R (R Core Team, 2017). A series of likelihood-ratio tests were used to determine if including the age-specific PGS and/or the common factor PGS as predictors improves model fit above a model with covariates only. The model with just covariates is referred to as the base model. Contrasts were drawn between the effect size for the age-specific PGS and the common factor PGS using a Z-test of the null hypothesis that two regression coefficients are equal (Paternoster et al. 1998).

Power Analysis

To address the third aim of this project, a simulation was conducted to determine the sample size required to achieve adequate statistical power for GWAS discovery in the adolescence developmental period under different assumptions regarding SNP MAF and effect size given the observed level of genetic correlation between developmental periods.

First, early adult and adult variables were simulated from a multivariate normal distribution with mean equal to 0 and correlation equal to the observed point estimate of the genetic correlation between early adult and adult alcohol use frequency using the mvtnorm package (Genz et al. 2021; Genz & Bretz, 2009) in R (R Core Team, 2017). A SNP variable was simulated as two draws from a binomial distribution to model a diploid genotype with probability equal to varying values for MAF (.01, .10, .20, .30, .40, .50). The adolescence variable was simulated from a normal distribution with the mean and variance conditioned on the early adult variable, the adult variable, and the SNP. The effects of the early adult variable and the adult variable on the adolescence variable were set to the corresponding observed genetic correlation point estimates. The effect of the SNP on the adolescence distribution was set to β=0.01. This effect size is comparable to those observed for significant SNPs in a recent GWAS of alcohol consumption (Liu et al. 2019). The variance of the adolescence variable was set to the square root of 1 minus the sum of squared effects on the adolescence indicator. Figure S12 displays a diagram of the data generation model.

These four indicators were used to construct a simulation model using the lavaan package (Rosseel, 2012) in R (R Core Team, 2017). The variance of the SNP was set using the formula for the variance of a SNP: 2MAF(1-MAF). A common factor was fit to the simulated adolescence, early adult, and adult variables by estimating all factor loadings and setting the variance of the common factor to 1. Residual variances for each of the simulated variables were freely estimated. The adolescence variable and the common factor were regressed on the SNP simultaneously. The simulation was repeated 1,000 times for each combination of the following parameters: MAF = .01, .10, .20, .30, .40, .50; SNP β=0.01, 0.05; N = 1k, 10k, 50k, 100k, 200k, 300k, 400k, 500k. Power for each combination of parameters was estimated at two different thresholds (p < 5e-8 and p < 1e-5) as the proportion of SNP p-values less than the threshold. The threshold p < 5e-8 represents the common criteria for genome-wide significance (Z. Chen et al. 2021; Risch & Merikangas, 1996), and the threshold p < 1e-5 represents the criteria for a “suggestive” effect (Lander & Kruglyak, 1995). Figure S13 displays a diagram of the simulation analysis model.

Results

Linkage Disequilibrium Score Regression (LDSC)

Univariate and cross-trait LD score regression intercepts, estimates of SNP-based heritability, genetic covariances, and genetic correlations are presented in Table 1. SNP-based heritability for each trait was modest (Adolescence H²_SNP = .04, SE=0.04; Early Adulthood H²_SNP = .05, SE = 0.05; Adulthood H²_SNP = .08, SE = 0.04). Genetic correlations between adolescence and early adulthood (rG = −.27, SE = 0.79), adolescence and adulthood (rG = −.34, SE = 0.65), and early adulthood and adulthood (rG = .75, SE = 0.58) had very wide confidence intervals and were not statistically significant. Notably, all estimates of H²_SNP were small, indicating that the genetic correlations reported here account for only a small proportion of the overall relationship between these phenotypes.

Table 1.

Summary of LD-Score Regression Results for Alcohol Use Frequency

LD-Score Regression Intercepts
	Adolescence	Early Adult	Adult
Adolescence	1.00	0.15	0.02
Early Adult	0.15	1.00	0.17
Adult	0.02	0.17	1.00
Heritability and Genetic Covariance
	Adolescence	Early Adult	Adult
Adolescence	.04 (0.04)	−.01 (0.03)	−.02 (0.03)
Early Adult	−.01 (0.03)	.05 (0.05)	.05 (0.03)
Adult	−.02 (0.03)	.05 (0.03)	.08 (0.04)
Genetic Correlations
	Adolescence	Early Adult	Adult
Adolescence	1	−.27 (0.79)	−.34 (0.65)
Early Adult	−.27 (0.79)	1	.75 (0.58)
Adult	−.34 (0.65)	.75 (0.58)	1

Note. Standard errors for parameter estimates are provided in parentheses.

Genomic Structural Equation Model GWAS

Loadings of genetic variance in each developmental period on the common factor were: adolescence λ_12-17 = .35 (SE = 0.64)_, early adulthood λ_18-25 = −.78 (SE = 1.12)_, adulthood λ₂₆₊ = −.97 (SE = 1.36). Residual genetic variances in each developmental period after accounting for the common factor were: adolescence U_12-17 = .88 (SE = 1.21)_, early adulthood U_18-25 = .40 (SE = 2.11)_, adulthood U₂₆₊ = .07 (SE = 2.61).

Manhattan plots and quantile-quantile plots depicting SNP effects on the components of the model are included in Figure 3 and Figure 4. In adolescence, three SNPs met the threshold for genome-wide significance of p<5e-8 (rs116734991, rs115778926, rs117048287). Note that these SNPs have not been detected in previous, larger GWAS of alcohol use outcomes. An additional 32 SNPs met the suggestive threshold of p<1e-5 in adolescence. No SNPs reached genome-wide significance in the other GWAS. In early adulthood, 67 SNPs met the suggestive threshold. In adulthood, 38 SNPs met the suggestive threshold. In the common factor GWAS, 30 SNPs met the suggestive threshold. Lists of all SNPs meeting these thresholds in adolescence, early adulthood, adulthood and the common factor are available in Table S4, Table S5, Table S6, and Table S7.

Figure 3.

Manhattan plots of gSEM GWAS of total genetic variance in alcohol use frequency in the full sample (ALSPAC, COGA, and Add Health).

The genome-wide significance (p<5e-8) is marked in red. The threshold for suggestive significance (p<1e-5) is marked in blue.

Figure 4.

Quantile-quantile plots of gSEM GWAS of total genetic variance in alcohol use frequency in the full sample (ALSPAC, COGA, and Add Health).

Observed −log10(p) are plotted against the expected uniform distribution of −log10(p) under the null hypothesis.

Genetic Correlations with External GWAS

Genetic correlations between gSEM GWAS summary statistics and the external GWAS are presented in Figure 5. No significant associations were observed with alcohol use frequency in adolescence. Drinks per week (rG=.80, 95% CI [0.06, 1.54]), alcohol use frequency (rG=.89, 95% CI [0.02,1.76]), and educational attainment (rG=.40, 95% CI [0.01, 0.80]) were significantly associated with alcohol use frequency in early adulthood. Drinks per week (rG=.84, 95% CI [0.36, 1.31]), AUDIT-P (rG=.54, 95% CI [0.15, 0.93]), alcohol use frequency (rG=.73, 95% CI [.29, 1.17]), alcohol use quantity (rG=.58, 95% CI [0.18, 0.98]), binge drinking frequency (rG=.83, 95% CI [0.33, 1.32]), and a common factor of externalizing behaviors (rG=.23, 95% CI [0.03, 0.43]) were significantly associated with alcohol use frequency in adulthood. Drinks per week (rG=.88, 95% CI [0.47, 1.28]), AUDIT-P (rG=.50, 95% CI [0.19, 0.82]), alcohol use frequency (rG=.89, 95% CI [0.46, 1.31]), alcohol use quantity (rG=.45, 95% CI [0.14, 0.76]), binge drinking frequency (rG=.79, 95% CI [0.39, 1.19]), cigarettes per day (rG=−.27, 95% CI [−0.51, −0.03]) and educational attainment (rG=.26, 95% CI [0.08, 0.45]) were significantly associated with the alcohol use frequency common factor.

Figure 5.

Genetic correlations between alcohol use frequency at different developmental periods and a series of other phenotypes in the full sample (ALSPAC, COGA, and Add Health).

Adolescence, Early Adulthood, and Adulthood refer to the Genomic Structural Equation Model GWAS results using the total variance parameterization of the model The X-axis of the plot is truncated at −1 and 1. A list of the phenotypes with sample sizes and citations is included in Table S3. Abbreviation: AUDIT-P = Alcohol Use Disorder Identification Test, Problems Subscale.

Polygenic Scores (PGS)

Univariate and cross-trait LD score regression intercepts, estimates of SNP-based heritability, genetic covariances, and genetic correlations for the LD score regression analysis with COGA removed are presented in Table S8. Manhattan plots and quantile-quantile plots depicting SNP effects on the residual components of the model are included in Figures S4 - S11. Descriptive statistics for phenotypic measures in the COGA analytic sample can be found in Table S2. Briefly here, the sample sizes in COGA for PGS analyses were n=1,118 in adolescence, n=2,762 in early adulthood, and n=5,255 in adulthood. Average drinking days per year in COGA were M=36.15 (SE=1.60) in adolescence, M=96.45 (SE=1.75) in early adulthood, and M=98.41 (SE=1.67) in adulthood.

In adulthood, including the common factor PGS as a predictor improved model fit relative to a model with just covariates (χ2(1)=4.36, p=.037). Including the age-specific PGS as a predictor also improved model fit relative to a model with just covariates (χ2(1)=6.43, p=.011). Including the age-specific PGS as a predictor did not improve model fit relative to a model with covariates and the common factor PGS (χ2(1)=2.89, p=.089). Likelihood-ratio tests comparing model fit between versions of the model with and without the residual and common factor PGS are included in Table 2. Additional model fit statistics are provided in Table S9.

Table 2.

Likelihood-Ratio Tests for PGS Linear Mixed-Effects Models in COGA

		Chi Square	df	p
Adolescence
	Base VS Common Factor PGS	0.04	1	.836
	Base VS Age-specific PGS	0.41	1	.520
	Common Factor PGS VS Both PGS	0.38	1	.539
Early Adult
	Base VS Common Factor PGS	3.07	1	.080
	Base VS Age-specific PGS	0.00	1	.956
	Common Factor PGS VS Both PGS	0.03	1	.860
Adult
	Base VS Common Factor PGS	4.36	1	.037
	Base VS Age-specific PGS	6.43	1	.011
	Common Factor PGS VS Both PGS	2.89	1	.089

Note. The Base model includes only covariates (sex, birth year, 10 ancestry PCs, and individual proband status in early adulthood and adulthood). All models include covariates. Abbreviation: PGS = Polygenic Score

The age-specific PGS (B=3.14, SE=2.82, 95% CI [−2.40, 8.67]) and common factor PGS (B=1.69, SE=2.78 , 95% CI [−3.76, 7.13]) were not associated with alcohol use frequency in adulthood when modeled together. Regression coefficient estimates for the age-specific PGS, the common factor PGS, and covariates at each developmental period are included in Table 3. All regression coefficient estimates are presented in their original scale of drinking days per year. The z-test for the null hypothesis that the age-specific PGS B and common factor PGS B are equal was not significant (z=0.45, p=.650).

Table 3.

Regression Coefficient Estimates from PGS Linear Mixed-Effects Models in COGA

	B	SE	P	Lower 95%	Upper 95%
Adolescence
Age-specific PGS	−1.03	4.23	.808	−9.31	7.25
Common Factor PGS	−0.14	1.60	.932	−3.28	3.00
Birth Year	−1.10	0.27	<.001	−1.63	−0.58
Female	−8.73	3.17	.006	−14.95	−2.51
Early Adult
Age-specific PGS	−0.31	1.75	.860	−3.74	3.13
Common Factor PGS	3.06	2.58	.235	−1.99	8.11
Birth Year	0.57	0.23	.013	0.12	1.03
Female	−38.18	3.39	<.001	−44.82	−31.55
Proband	29.95	8.42	<.001	13.46	46.45
Adult
Age-specific PGS	3.14	2.82	.267	−2.40	8.67
Common Factor PGS	1.69	2.78	.544	−3.76	7.13
Birth Year	0.71	0.09	<.001	0.53	0.90
Female	−55.02	3.06	<.001	−61.02	−49.02
Proband	95.95	5.03	<.001	86.08	105.82

Note. Reported standard errors, p-values, and 95% confidence intervals are corrected for false discovery rate. Abbreviations: PGS = Polygenic Score; B = Unstandardized regression coefficient; SE = Standard Error.

The age-specific PGS was associated with adult alcohol use frequency when modeled separately from the common factor PGS (B=4.00, SE=1.89, 95% CI [0.31, 7.70]. The association between the common factor PGS and adult alcohol use frequency was not significant when modeled separately from the age-specific PGS (B=3.31, SE=1.59, 95% CI [−0.75, 7.38]. Regression coefficient estimates from analyses with each PGS modeled separately from each other in each developmental period are available in Table S11 and Table S12. Note that the variance accounted for by the age-specific PGS in adulthood (0.1%) is substantially smaller than the variance accounted for by PGS from larger, developmentally-agnostic GWAS (2.4%; Liu et al. 2019). Change pseudo-R² values for the residual and common factor PGS are reported in Table S10.

The age-specific PGS and common factor PGS were not significantly associated with alcohol use in adolescence and early adulthood. The PGS did not improve model fit when tested by LRT (Table 2), accounted for extremely small amounts of phenotypic variance (Table S10), and were not significantly associated with the target phenotype (Table 3). The effect of the age-specific PGS were not significantly larger than the effect of the common factor PGS (adolescence z=−0.35, p=.725; =−1.32, p=.187 early adulthood).

Power Analysis

Manifest indicators were simulated using the observed genetic correlations between adolescence, early adulthood, and adulthood (rG_Adol,E.Adult = −.34, rG_Adol,Adult = −.27, rG_{E.Adult,Adult}=.75).

For the genome-wide significance threshold (p < 5e-8), at least 80% power to detect a SNP with effect size β=0.01 was not achieved under any combination of simulation parameters. For p < 5e-8, MAF=.5 and n=500,000 (the most well-powered combination of parameters) power to detect a SNP with effect size β=0.01 was 9%. For the suggestive threshold (p < 1e-5), at least 80% power to detect a SNP with effect size β=0.01 was not achieved under any combination of simulation parameters. For p < 1e-5, MAF=.50 and N=500,000 power to detect a SNP with effect size β=0.01 was 42% Power estimates for genome-wide significance (p < 5e-8) and suggestive significance (p < 1e-5) at varying values of MAF are presented for a SNP with effect size β=0.01 in Figure 6.

Figure 6.

Power estimates for the adolescence residual component of the gSEM at the genome-wide significance threshold (p < 5e-8; left) and the suggestive threshold (p<1e-5; right) with SNP effect size β=0.01

Vertical lines in the figure demarcate the observed sample sizes in the discovery analysis of the current study: dotted = adolescence, dot-dashed=early adulthood, dashed = adulthood, solid = common factor.

Discussion

This work addressed three aims: (1) to advance gene discovery by building developmentally-informative models for gene identification capable of incorporating developmental changes in alcohol use across time (2) to leverage results from the developmentally-informative GWAS for genetic prediction of age-matched alcohol use outcomes in an independent sample and (3) to estimate the required sample size for an adequately powered implementation of the model.

Genetic Correlations Across Development

The 95% confidence interval for the genetic correlation between alcohol use frequency in adolescence and adulthood was distinct from one (rG= −.34, 95% CI [−1.61, 0.93]), suggesting that there is heterogeneity in the genetic liability that underlies alcohol use frequency throughout development. Notably, all estimates of H²_SNP were small and non-significant (Adolescence H²_SNP = .04, SE=0.04; Early Adulthood H²_SNP = .05, SE = 0.05; Adulthood H²_SNP = .08, SE = 0.04), indicating that the genetic correlations reported here account for only a small proportion of the overall relationship between these measures. Small estimates of H²_SNP and the large confidence intervals for each genetic correlation indicate that these results should be interpreted with caution.

Genetic Correlations with External GWAS

Genetic correlations between the gSEM components and external GWAS provide a basis to begin disentangling what comprises differences in genetic liability across development. Previous work demonstrates that the genetic liability underlying alcohol use frequency in adulthood is different than the genetic liability underlying other alcohol use outcomes in adulthood, such as alcohol problems and alcohol use quantity (Walters et al. 2018; Sanchez-Roige et al. 2019; Kranzler et al. 2019; Mallard et al. 2022b). Negative genetic correlations are reported between adult alcohol use frequency and externalizing phenotypes, while positive genetic correlations are reported between adult alcohol use frequency and indicators of socioeconomic status, such as educational attainment and income (Mallard et al. 2022b). The opposite pattern of results is observed for adult alcohol problems and alcohol use quantity: positive genetic correlations with externalizing phenotypes and negative genetic correlations with indicators of socioeconomic status (Mallard et al. 2022b). These previous results suggest that adult alcohol use frequency measures may index a variety of socioeconomic and environmental constructs that are not of immediate relevance in describing the etiology of clinically relevant alcohol use behaviors (Kranzler et al. 2022; Mallard et al. 2022b). In the current study, adult alcohol use frequency demonstrated nominally a negative genetic correlation with cigarettes per day and a nominally positive genetic correlation with educational attainment. Though these results were not statistically significant, these point estimates align with previous work (Kranzler et al. 2022; Mallard et al. 2022b) to suggest that adult alcohol use frequency is dissimilar from other adult alcohol use behaviors that demonstrate association with externalizing outcomes.

It is possible that the negative genetic correlation between alcohol use frequency and externalizing phenotypes does not extend to adolescence. In adolescence, the point estimates of genetic correlations with educational attainment and average household income were near zero, departing from the patterns seen in adulthood. Additionally, the point estimate of the genetic correlation between adolescent alcohol use frequency and risk tolerance (rG=.35, 95% CI [−0.07, 0.77]) suggests that the genetic underpinnings of adolescent alcohol use frequency are more closely related to externalizing behavior, though none of the genetic correlations between phenotypes from external GWAS and alcohol use frequency in adolescence were significantly different than zero. These suggestive results align with previous findings that externalizing genetic risk factors are especially important in adolescence (Kendler et al. 2011; Meyers et al. 2014). The same behavior measured at different ages may have different genetic architectures as well as other complex behavioral environmental causes, correlates, and consequences depending on the developmental context of the behavior. For example, whether alcohol can be obtained legally (Wagenaar and Toomey, 2015) and parental monitoring (Dick et al. 2009; Kendler et al. 2011) are important factors in shaping the context of alcohol use and, in turn, the genetic factors that are associated with alcohol use.

Age-Matched PGS Validation

In adulthood, the age-specific PGS and the common factor PGS each improved model fit above a model with just covariates. The regression coefficient for the age-specific PGS was also significant after correcting for multiple testing when the common factor PGS was excluded from the model. The significant regression coefficient associated with the age-specific PGS predicted an increase of approximately 4 drinking days per year per PGS standard deviation. The common factor PGS was not significantly associated with alcohol use frequency in adulthood after correcting for multiple testing. Nominal differences in AIC align with these results, indicating that the model with the age-specific PGS and covariates provided the best fit to the data.

While these results provide tentative evidence for the utility of the age-specific PGS in adulthood, the observed effect of the age-specific PGS was relatively small. Nominal differences in BIC also depart from this pattern, instead suggesting that the model with just covariates provided the best fit to the data. Neither PGS demonstrated significant effects when modeled together and the age-specific PGS did not improve model fit significantly above a model that included the common factor PGS when tested via likelihood-ratio test. Larger samples may be required to more clearly delineate the effects of the common factor PGS and the age-specific PGS. The results of the power simulation suggest that future developmentally-informative GWAS that apply the approach described in the current study may need large sample sizes, in excess of 500,000, to detect reliable SNP effects.

Limitations

The results of this study should be interpreted in the context of several important limitations. Foremost, the GWAS meta-analysis that was used to generate summary statistics within developmental periods was underpowered. Sample sizes were 8,869.29, 9,647.64, and 9,894.18 in adolescence, early adulthood, and adulthood, respectively. COGA was removed from the discovery analysis in Aim 2 to facilitate PGS construction, further reducing power. Second, estimates of H²_SNP were small in each developmental period and the 95% confidence intervals for estimates derived from all stages of the analysis pipeline were large. As a result, point estimates should be interpreted cautiously. Third, these analyses only included European ancestry participants, limiting the generalizability these findings across ancestry groups. It is a critical priority to extend genetic analyses to a broader range of ancestry groups (Peterson et al, 2019). Fourth, we did not apply any transformations to our phenotypes, despite right-skew and zero-inflation in the distribution of adolescent alcohol use. This allows for intuitive interpretation of effect sizes in units of drinking days per year, but may have impacted results. Analysis with log transformed phenotypes (ln(phenotype+1)) produced negative heritability estimates in early adulthood (H²_{SNP =}−0.0041 SE=0.0452), preventing complete assessment of this approach. The genetic correlations between the transformed and untransformed versions of adolescence (rG=1.00, SE=.21) and adulthood (rG=.96, SE=.16) suggest that the pattern of results would be similar with log transformed phenotypes. The genetic correlation between log transformed adolescence and log transformed adulthood was small and positive with a large standard error (rG=.14, SE=.71), suggesting that we do not have adequate power to determine the direction of this association. Fifth, our power analysis was not constructed to account for the phenotypic distribution of our data, instead focusing on the genetic correlations between phenotypes. Accuracy of our power estimates may vary as a function of phenotype distribution. Finally, these analyses model age as a series of ordinal developmental periods (adolescence age 12-17, early adulthood age 18-25, adulthood age 26+), rather than a continuum. This approach is not sensitive to differences in genetic liabilities that may exist within the specified developmental periods; for example, the 18-25 age group includes subjects who cannot drink legally (< age 21) and others who can drink legally (> age 21) in Add Health and COGA. The legal drinking age in the U.K. (ALSPAC) is 18, further increasing heterogeneity in this analysis. Modeling age as a continuum, rather than discrete groups, may identify additional differences in genetic liabilities within these age groups.

Conclusions

Genetic influences on alcohol use vary throughout development (Aliev et al. 2015; Dick et al. 2006; Edwards & Kendler, 2013; Kendler et al. 2011; Meyers et al. 2014; Sakai et al. 2010). The omission of developmental considerations from gene-identification studies for alcohol use behaviors limits the utility of polygenic scores for phenotype prediction across the lifespan (Elam et al. 2021; Kandaswamy et al. 2021). Recent advances in multivariate genomic methods provide an opportunity to identify genetic variants that account for developmental changes in genetic liability for alcohol use behaviors. In this work, we present an analytic approach for developmentally-informative gene-identification using Genomic SEM. The age-specific PGS predicted an increase of 4 drinking days per year per PGS standard deviation when modeled separately from the common factor PGS in adulthood. Though the discovery analysis in the current work was underpowered, developmentally-informed gene discovery analyses may improve phenotype prediction via polygenic scores when discovery samples are adequately large to model differences in genetic liabilities across development. Beyond applications to polygenic scoring, larger developmentally-informative discovery analyses may also facilitate the identification of age-matched genetic instruments for improved Mendelian randomization studies. Our results suggest that very large sample sizes will be required to reach this goal using the approach described in the current work. The results presented here are an initial step towards toward this goal and lay a foundation for future molecular genetic studies of developmental variability in the genetic underpinnings of alcohol use behaviors and the subsequent possibility of genetically-informed, age-matched phenotype prediction.