Unraveling time-dependent genetic components underlying alcohol response

Abstract

While numerous studies have examined the subjective response to alcohol as an intermediate phenotype to understand its variability, heritability, and predictive capacity for alcohol-related disorders, in-depth analyses linking alcohol reactivity indicators to genetic factors within a large cohort have been absent. Our study aimed to quantify the exact contribution of each genetic variant relevant to the alcohol metabolism to the variability in alcohol response. Specifically, we focused on two primary genes involved in alcohol metabolism (ALDH2 and ADH1B) and three additional loci (ALDH1B1, ALDH1A1, and GCKR) that have been shown to have significant associations with drinking behaviors in Japanese individuals. We conducted the first study to assess the relationship between subjective response to alcohol (SR), evaluated by various assessment subscales, and genetic factors using an intravenous clamp technique in 429 healthy Japanese young adults. By reducing the dimensionality of the data to assess similarity structures, we identified three distinct clusters of SRs and participants. Each participant cluster exhibited a distinct alcohol response profile shaped by specific genetic contributions. Participant cluster 1 demonstrated the strongest response, followed by participant cluster 2, and then participant cluster 3. Participant cluster 1 may also be the most strongly influenced by the allelic status of ALDH2 and ADH1B. SR patterns varied accordingly, and the enrichment of the ALDH2*2 and ADH1B*2, differed across both participant and subscale clusters. Notably, the three participant clusters closely aligned with the three subscale clusters, highlighting a consistent genotype–phenotype relationship. Furthermore, the proportion of variance explained by these genes also varied across subscale clusters. Contrary to known functions, ADH1B showed associations at later timings when ALDH2 associations attenuate. Our three-cluster classification may improve prevention by enabling early identification of individuals at health risk.

Introduction

In 2016, harmful alcohol use led to three million deaths worldwide, 5.3% of the total, as reported by the WHO’s 2018 Global Status Report on Alcohol and Health [1]. Alcohol metabolism varies due to environmental and genetic factors. The two primary genes involved in alcohol metabolism are alcohol dehydrogenase 1B (ADH1B) and aldehyde dehydrogenase 2 (ALDH2, eFig. 1) [2]. Variants of these genes are population-specific. Rs1229984 (C > T) is a missense variant in ADH1B, with the T allele more common in East Asians compared to Europeans (70% vs 3%, respectively) [3]. Carriers of the rs1229984 T allele, also known as ADH1B*2, have an elevated activity level in metabolizing alcohol to acetaldehyde [4–6]. Similarly, rs671 (G > A), a missense variant in ALDH2, is prevalent in East Asians (17%) but is not found in Europeans [3]. Carriers of the rs671 A allele, known as ALDH2*2, have a lower activity level in metabolizing acetaldehyde to acetate [4, 5, 7]. However, it is noted that these two genes alone cannot fully explain the alcohol-related behavioral changes. Matoba et al. recently identified three additional loci, rs1260326 in GCKR, rs3043 in ALDH1B1, and rs8187929 in ALDH1A1, showing significant associations in a genome-wide association study (GWAS) for drinking behaviors in Japanese populations [8]. The subjective response to alcohol (SR) serves as an endophenotype to understand alcohol’s pharmacological impact (eTable 1). Research shows SR’s significant variability, heritability, and predictive power for alcohol-related disorders [9, 10]. However, comprehensive genetic analyses of SR in large cohorts have been limited. Therefore, this study aims to address this gap by unraveling the time-dependent genetic components underlying alcohol response.

Materials and methods

Study participants

The study enrolled 429 healthy Japanese young adults (volunteers) aged 20–24 years (eTable 2) from the National Hospital Organization Kurihama Medical and Addiction Center (Yokosuka, Japan) up until November 2020. All participants were in good health with no known physical or psychiatric illnesses, including alcohol use disorder. However, given the nature of our study involving alcohol administration, we specifically included individuals with a history of drinking. Consequently, it was anticipated that individuals with the ALDH2*2/*2 genotype would not participate in this study, as they cannot tolerate alcohol, and therefore, the frequency of this genotype would be lower than in the general population. Participant eligibility was determined through health status questionnaires and a review of past medical records. The exclusion criteria included individuals with medical histories of cardiovascular, pulmonary, renal, hepatic, or gastrointestinal diseases, or a personal history of any DSM-IV (American Psychiatric Association, 1994) axis I disorder, including alcohol-related disorders. Of note, although this project is an entirely independent study and all participants were recruited and studied specifically for this research, the previous study by Ueno et al. [11] was conducted using samples that partially overlap with this study. However, Ueno et al. focused on investigating blood acetaldehyde and alcohol concentrations in relation to cognitive function, without conducting whole genome typing, SR assessments using multiple tools, or other analyses as described in the methods section of this study. Therefore, our research and their research are distinct and separate investigations. The study’s protocol was approved by the ethical committees of the National Hospital Organization Kurihama Alcoholism Center (Yokosuka, Japan), the RIKEN Center for Integrative Medical Sciences (Yokohama, Japan), and the National Center for Geriatrics and Gerontology (Aichi, Japan). All participants provided written informed consent. Procedures contributing to this work comply with relevant ethical standards.

Genotyping of ADH1B and ALDH2 and whole-genome genotyping

DNA was extracted using the DNA Extractor WB Kit (Wako Pure Chemical Industries, Ltd., Osaka, Japan). The genotypes of ADH1B and ALDH2 were determined using polymerase chain reaction–restriction fragment length polymorphism (PCR-RFLP) methods [12, 13]. Whole-genome genotyping was conducted using the Infinium Asian Screening Array (ASA) v1.0 BeadChip.

Quality control (QC) of whole -genome genotyping data

QC for the post-whole genome genotyping samples and single nucleotide polymorphisms (SNPs) were performed. Refer to eMethods for details. In total, five samples were removed during the QC process.

Whole-genome imputation

Whole-genome imputation utilized a recently developed reference panel [14]. Refer to Flanagan et al. and eMethods for reference panel details [15]. This study specifically examined five variants, including rs671 in ALDH2 and rs1229984 in ADH1B, along with three other loci: rs1260326 in GCKR, rs3043 in ALDH1B1, and rs8187929 in ALDH1A, which was associated with drinking behaviors in the Japanese population [8].

In this study, we adopted two genetic analysis approaches: candidate analyses and genome-wide association studies (GWAS). We utilized the GWAS results for polygenic predictions.

Evaluation of alcohol response

The Intravenous (IV) clamp technique was performed [16]. Refer to the eMethods for details. Briefly, the Indiana BrAC clamp method was used to infuse a 6% (v/v) ethanol solution, targeting a BrAC level of 50 mg%. The target BrAC level was achieved within 15 min, followed by a 165-min infusion period to maintain this level. Afterward, participants were monitored until their BrAC decreased below 20 mg/dL. During the infusion, participants completed assessments with three SR assessment tools at steady alcohol concentrations every 30 min, as listed in eTable 1: 1. The Bodily Sensation Scale (BSS), which measures the subjective experience of physiological changes and comprises six components, including anesthetic, central stimulant, impaired function, warmth, dynamic peripheral, and general sensations [17]. 2. The Biphasic Alcohol Effects Scale (BAES), which evaluates sedative and stimulant effects. We used the original form of the BAES, with separate components for sedation and stimulation [18]. 3. The Subjective High Assessment Scale (SHAS), which assesses perceptions about various sensations often associated with alcohol but not specifically attributed to it, consisting of three components: intensity of high (SHAS1), degree of drunkenness (SHAS2), and sleepiness (SHAS3). This tool is a modification of the SHAS described by Schuckit et al. in Arch Gen Psychiatry 1988 [19]. The SHAS score is based on a visual analog scale, where participants marked a position on a continuous line. The distance from the leftmost point was measured and divided by the total length of the scale to obtain a score, which was then used for analysis. This study employed three distinct assessment tools (BSS, BAES, and SHAS), each termed a scale, to facilitate a comprehensive evaluation of SR. Each scale comprised specific categories, termed subscales (The SHAS subscales are, in fact, single items). Assessments began with baseline measurements right before the infusion, followed by subsequent evaluations every 30 min up to 240 min, and at two points after the infusion ceased. Participants assigned scores to these subscales within each scale, using these scores as a quantitative method to assess their SR. The analysis primarily focused on scores at the 30-min and 150-min intervals, chosen due to the observation that SR typically peaks at these times during alcohol administration. The rationale and detailed findings of this focused analysis are elaborated in the Results section.

Comparisons of scores from the subscales of the BSS, BAES, and SHAS scales, based on the allele differences of ALDH2 and ADH1B, as well as genotypes of ALDH1A1, ALDH1B1 and GCKR, were initially performed using two-way repeated measures analysis of variance (ANOVA), with time as the within-group factor and allele differences or genotypes as the between-subject factors. We also confirmed these results using an ANCOVA by adding covariates (PC1–5), age, sex, current weekly alcohol consumption (based on drinking days per week), or habitual drinking, or family history of alcohol-use disorders (AUD) to ensure that the findings did not change after adjusting for these variables. Of note, 47 participants (11%) had a family history of AUD in this study. In addition, t tests and Kruskal-Wallis tests were subsequently performed for comparisons between two groups and among three or more groups, respectively. Analyses were two-sided, with a P value of <0.05 indicating significance. For the comparisons of subscale scores across participant clusters, we performed one-way ANOVA. Additionally, to evaluate genotype-related effects on alcohol response at the subscale cluster level, we calculated z-scores from the p-values obtained by comparing the effects of ALDH2 and ADH1B alleles at the 30- and 150-min time points for each subscale cluster, using a two-tailed transformation. To prevent computational infinity from extremely small p-values, z-scores were capped at a maximum of 8.0. Multiple-testing correction was applied for analyses with multiple comparisons. In this study, statistical analyses were conducted with R version 4.0.2, unless otherwise noted throughout the manuscript [20].

Normalization of subscales (BSS, BAES and SHAS)

We set scores at 0-min point as reference and subtracted them from scores at each time point, followed by normalization of the values to mean 0 and variance 1 at each time point (see eMethods for details). Samples lacking data at the 0-min point were excluded from the analysis of the alcohol clamp technique results. Specifically, exclusions included: 17 samples for BSS (anesthetic), 8 samples for BSS (central stimulant), 9 samples for BSS (impaired function), 7 samples for BSS (warmth), 8 samples for BSS (dynamic peripheral), 20 samples for BSS (general), 6 samples for BAES (sedative), 7 samples for BAES (stimulant), and 4 samples for all three SHAS scales. Following this, the Shapiro-Wilk test was used to evaluate the normality of the scores from BSS, BAES, and SHAS. Given the histograms indicated an abnormal distribution, rank-based inverse normal transformations were applied to each participant’s score from BSS, BAES, and SHAS to achieve a normal distribution.

Clustering of subscales and participants

To determine the underlying shared dimensions of the three assessment scales (BSS, BAES, and SHAS), we conducted a cluster analysis. This analysis was based on the scores from each subscale at the 30-min and 150-min points, allowing for the observation of shifts in relationships over time. Refer to the eMethods for details.

Principal component analysis (PCA)

We explored the separation of subscales within the BSS, BAES, and SHAS scales based on the scores for each participant through PCA. We conducted PCA on the participants using the scores at the 30- and 150-min points. Refer to the eMethods for details.

Quantitative trait analysis to investigate additivity of ALDH2 and ADH1B

We performed a quantitative trait analysis to examine the additivity of ALDH2 and ADH1B at the 30-min point. This was done employing the dominant model for ADH1B and based on possession of *2 for ALDH2. Initially, we implemented a linear regression model with normalized scores of BSS, BAES, and SHAS as dependent variables and the allelic status of either ALDH2 or ADH1B as explanatory variables. Then, we applied a linear regression model with the normalized scores of BSS, BAES, and SHAS as dependent variables, and the allelic status of both ALDH2 and ADH1B as explanatory variables. Lastly, we used a linear regression model with the normalized scores of BSS, BAES, and SHAS as dependent variables, the interaction between ALDH2 and ADH1B as explanatory variables and the allelic status of ALDH2 combined with ADH1B as covariates.

Investigation of the relationship between participants’ PCs and subscales

We examined relationships between participants’ principal components (PC1 and PC2) and various subscales at 30 and 150 min intervals. After PCA, we used the derived PCs as predictor variables in linear regression models for each subscale.

Comparison of effect sizes of the lead single nucleotide polymorphism (SNP) in the ALDH2 and ADH1B

We calculated the effect sizes using linear regression analyses, with the normalized scores of BSS (warmth) as the dependent variable and the allelic statuses of ALDH2 and ADH1B as the explanatory variables.

Assessment of acute adaptation

Based on previous studies by Roh et al. [21] and Morzorati et al. [22], we evaluated the presence of acute adaptation using repeated-measures ANOVA. Given that we maintained the target BrAC level of 50 mg% with a 165-min infusion period, we defined the dependent variables as the changes in SR measurements from 30 to 180 min after the infusion, measured every 30 min for a total of six measurements.

Variance component analyses

We assessed total variance in BSS, BAES, and SHAS scores through a linear regression model with tag SNPs of ALDH2, ADH1B, ALDH1B1, ALDH1A1, and GCKR.

Variance partitioning analysis

Using the R package variancePartition (v1.28.0), we sought to understand the source of variation in SRs [23]. We considered tag SNPs from the aforementioned genes, along with age, sex, and BMI as variables. This analysis employed a linear mixed model to quantify the variation in SR attributable to these variables.

Polygenic risk score (PRS)

We performed GWAS to assess PRS, using scores from BSS, BAES, and SHAS subscales as phenotypes. After excluding regions reported by a previous study, PRS was constructed to evaluate the polygenic effects of unknown regions (see eMethods for details).

Results

The study design is depicted in eFig. 2. The baseline characteristics of the participants are displayed in eTable 2 and eTable 3. As stated in the Methods section, this study evaluated SR through three assessment tools: the BSS, BAES, and SHAS scales. Each of these scales is comprised of subcategories, referred to as subscales. Participants assigned scores to each subscale to examine SR. Throughout this manuscript, the abbreviation PC consistently refers to principal component.

Characteristics of scores for each SR

The overall trend of scores for each subscale of the BSS, BAES, and SHAS scales, recorded every 30 min after alcohol administration, is depicted in eFig. 3. Given the variation in scores within the same scale, analysis was performed for each subscale. Peaks observed in each subscale occurred at slightly different time points, with the subsequent trends after these peaks varying (refer to eFig. 3). Most subscales (9 out of 11) reached their peak scores 30 min after administration, while the remaining two subscales peaked at 150 min. Additionally, the subscale data at 30 and 150 min were particularly informative for approximating the shapes of the time-dependent subscales. Following alcohol infusion, the target BrAC level of 50 mg% was maintained, which resulted in scores also achieving a plateau after administration.

Based on the peak overall trends in the subscale scores of the BSS, BAES, and SHAS, 30 and 150 min were determined as representative time points.

Three clusters of subscales and participants

To better capture similarities among subscales and reduce dimensionality —allowing us to identify underlying commonalities and independent components beyond surface-level patterns—we applied both cluster analysis and PCA. While BSS, BAES, and SHAS contain similar subscales, they represent distinct metrics, and each subscale alone does not provide a comprehensive understanding of the tendencies observed in the Japanese cohort. Therefore, we employed clustering to classify the subscales and used PCA to validate these classifications, ensuring that the identified clusters are supported by multiple analytical approaches. This approach allowed us to uncover shared characteristics, patterns of similarity, and distinct features across the subscales beyond their predefined categories. Employing a hierarchical clustering approach with a Euclidean distance matrix, we discerned three distinct clusters within the subscales of BSS, BAES, and SHAS (Fig. 1a). Subscale cluster 1 includes all subscales except for BAES (stimulant, subscale cluster 2), SHAS1 (subscale cluster 2), and SHAS3 (subscale cluster 3). These three subscale clusters ultimately demonstrated the most accurate classification of the data, a finding supported by subsequent analyses. Consistent with the initial subscale clustering results, employing either the alternative Pearson correlation coefficient or the alternative complete linkage method yielded identical subscale cluster configurations, resulting in the same three subscale clusters (Fig. 1b, c). To gauge the significance of these subscale clusters from an explained variance standpoint, we executed 1,000 permutations of the subscales to disrupt their covariance structure. None of these permutations achieved a similar variance explained by the three subscale clusters (Fig. 1d). Furthermore, we applied Gap statistics for k-means clustering, which reconfirmed that the subscales were best classified into three subscale clusters (Fig. 1e). All these findings consistently endorse the three subscale clusters. To further validate this classification, we performed PCA, which also supported these three subscale clusters (eFig. 4). This three-cluster classification was subsequently used for further analyses. Importantly, these subscale clusters remained largely consistent across various time points (eFig. 5). PC1 differentiated subscale cluster 1 from subscale clusters 2 and 3. The proportions of variance explained by PC1 and PC2 were 0.52 and 0.24, and 0.56 and 0.18 at 30 and 150 mins, respectively (eFig. 4), indicating that around 75% of the variance could be attributed to the top 2 PCs.

Fig. 1. Cluster subgrouping of participants into three clusters.

a Hierarchical clustering using Ward’s method. b Hierarchical clustering using complete linkage. c Heatmap of Pearson correlation coefficients for clusters. A stronger correlation is indicated by a more intense yellow, while a weaker correlation is shown in darker blue. d Distribution of the fraction of the sum of squares between three clusters over the total sum of squares in 1000 permutation tests. The statistic of the original three clusters is represented by a red line. e The K-means clustering with the highest Gap statistics is displayed. f Clustering of participants based on subscales.

Moreover, when clustering participants based on the subscales, we applied hierarchical clustering using Ward’s minimum variance method (ward.D2), with Euclidean distances computed between participants based on their subscale scores. This approach optimally minimizes within-cluster variance, ensuring that participants within the same cluster exhibit the most similar subscale score profiles (Fig. 1f). Based on this clustering, we identified three distinct participant groups. Subsequently,we established a correspondence between the three clusters of participants and the clustered subscales (see eFig. 6). Specifically, eFig. 6 illustrates that participants in cluster 1 (corresponding to the rightmost section of the horizontal axis) exhibited high scores in subscale cluster 1, where the highest color key values are found in the horizontal band corresponding to subscale cluster 1 on the vertical axis. Similarly, participants in cluster 2 (corresponding to the middle-right section of the horizontal axis) showed elevated scores in subscale cluster 3 (where the highest color key values in the horizontal band corresponding to subscale cluster 3 on the vertical axis), while participants in cluster 3 (corresponding to the leftmost section of the horizontal axis) exhibited higher scores in subscale cluster 2 (where the highest color key values are found in the horizontal band corresponding to subscale cluster 2 on the vertical axis). These findings further support the validity of the clustering results, demonstrating that both subscales and participants can be classified into three distinct groups beyond surface-level patterns. Moreover, the strong correspondence between participant clusters and subscale clusters suggests that, in the Japanese population, alcohol response can be systematically categorized into three biologically relevant groups. This represents a novel finding, providing new insights into how alcohol reactivity is structured within this cohort.

To further illustrate the relationship between subscale clusters and participant clusters, we additionally compared subscale scores across participant clusters at the 30-min time point, as shown in eTable 4. The results demonstrate that participant clusters 1, 2, and 3 exhibited distinct patterns of alcohol response, with statistically significant differences observed even after correction for multiple comparisons. Notably, participant cluster 1 showed the strongest response, followed by clusters 2 and 3 (participant cluster 1 > participant cluster 2 > participant cluster 3, based on the means shown in eTable 4). We also found that subscale cluster 1 exhibited the most pronounced differences among the three participant clusters, followed by subscale clusters 2 and 3 (subscale cluster 1 > subscale cluster 2 > subscale cluster 3, based on ANOVA p-values in eTable 4).

Attribution of genetic factors to scores of each SR

Next, we sought to determine the contribution of genetic variants in alcohol-metabolizing enzyme-encoding genes. First, we focused on the variants in the two major genes, namely, ALDH2 and ADH1B. Subscale scores were compared across subscale clusters based on the carrier status of alleles for these genes (eTable 3). When comparing scores in each cluster of the subscales between ALDH2 1*1 and ALDH2 1*2, distinct association patterns emerged among the three subscale clusters. Repeated measures ANOVA revealed that, overall, subscale cluster 1 demonstrated the strongest association with ALDH2, and this effect changed over time. In contrast, subscale cluster 2 showed the weakest association (eTable 5). When comparing scores at each time point using t tests (Fig. 2, eTable 6), those with the ALDH2 1*2 exhibited higher scores at multiple points in subscale cluster 1 as shown in Fig. 2, aligning with the known function of the ALDH2*2 allele, which reduces the activity of ALDH2. In subscale cluster 2, the association with the ALDH2*2 allele status was weaker than in subscale cluster 1 as suggested by fewer time points showing statistically significant differences in t-tests, particularly on the BAES stimulant scale. Notably, associations in subscale cluster 3 were only observed at later time points. This suggests that the three biologically relevant groups exhibited distinct functional effects of ALDH2, reflecting differences in its genetic influence.

Fig. 2. Comparisons of the scores of BSS, BAES and SHAS measured every 30 min after administering alcohol in the participants with ALDH2 1*1 vs ALDH2 1*2.

We compared scores of the three clusters, measured every 30 min after administering alcohol, between participants with ALDH2 1*1 vs ALDH2 1*2. The x-axis marks represent time points where score differences are significant (reaching a nominal p value of <0.05) between those with ALDH2 1*1 and ALDH2 1*2. Even after accounting for multiple testing (11 subscales across BSS, BAES, and SHAS, resulting in a cutoff of p = 0.0045), the significant trends remained almost unchanged, as shown in eTable 5. The x-axis represents the time after starting the alcohol infusion, while the y-axis denotes scores of each subscale. Notably, cluster classification is based on PCA results at the 30-min mark.

In the comparison of subscale scores across subscale clusters for ADH1B genotypes using t-tests, subscale cluster 2 displayed minimal or no association with ADH1B, as suggested by the number of time points showing statistically significant differences (Fig. 3, eTable 7). Once again, subscale cluster 3 displayed associations mainly at later time points, though these associations were only trends, as they did not reach the multiple-testing-adjusted significance threshold. This suggests that the three biologically relevant groups exhibited distinct functional effects of ADH1B, further highlighting differences in its genetic influence.

Fig. 3. Comparisons of the scores of BSS, BAES and SHAS measured every 30 min after administering alcohol in the participants with ADH1B 1*1 vs ADH1B 1*2 or 2*2.

We compared scores of the three clusters, measured every 30 min after administering alcohol, between participants with ADH1B 1*1 and either ADH1B 1*2 or 2*2. The x-axis points show the time points where the differences in scores are significantly different (reaching a nominal p value of <0.05) between these groups. Even after accounting for multiple testing (11 subscales across BSS, BAES, and SHAS, resulting in a cutoff of p = 0.0045), the significant trends remained almost unchanged, as shown in eTable 6. The x-axis represents the time point after starting the alcohol infusion, while the y-axis depicts scores of each subscale. The cluster classification is based on PCA results at the 30-min mark.

Our analysis of genotypes across the three clusters of participants revealed significant disparities in genotype frequencies. The frequency of the ADH1B*2 allele carriers was highest in participant cluster 2, decreasing in participant cluster 3, and least common in participant cluster 1 (p = 0.027, Fisher exact test), as depicted in eFig. 7 and eTable 8. Likewise, carriers of the ALDH2 *2 allele tended to be enriched in participant clusters 2 and 3. That is, our study population can be divided into three participant clusters, each characterized by distinct functional profiles of ALDH2 and ADH1B.

When considering the allelic patterns of both ALDH2 and ADH1B using repeated measures ANOVA, each of the three groups exhibited distinct involvement. Specifically, subscale cluster 1 exhibited the most pronounced association, with this effect changing over time (eTable 5). In contract, subscale cluster 2 displayed a weaker association, especially on the BAES stimulant scale. At specific time points, t-tests revealed that the groups with a combination of ALDH2 1*2 and either ADH1B 1*2 or 2*2 alleles exhibited significantly higher scales (eFig. 8 and eTable 9). This followed a trend similar to the subscale clusters observed in ALDH2 1*2 (Fig. 2 and eFig. 8, eTable 6). This observation aligns with the functional activities of ALDH2*2 and ADH1B*2. Regression analysis indicated that the two variants were independently associated with subscale scores (eTable 10). Therefore, we confirmed the additive effects of each allele on the subscales, adressing an under-investigated area highlighted in previous studies [24–26]. Furthermore, this analysis revealed a weaker association (smaller beta coefficients) in subscale clusters 2 and 3 compared to subscale cluster 1.

We also analyzed the contributions of three additional alcohol related genes: rs1260326 in GCKR, rs3043 in ALDH1B1, and rs8187929 in ALDH1A, which have been linked to drinking behaviors in the Japanese population (eTables 3, 11–13) [8]. Overall, the effects of GCKR, ALDH1A1, or ALDH1B1 also support the three-cluster classifications. That is, each gene exhibited a distinct pattern of involvement across the clusters, indicating that the three clusters differ in how these genes contribute to alcohol-related traits. When focusing on individual time points using t-tests, predominantly, only rs3043 in ALDH1B1 showed significant score differences, mainly in subscale cluster 1, as suggested by the number of time points reaching the multiple-testing-adjusted significance threshold (eFigs. 9–11). For the rs1260326 in GCKR and rs8187929 in ALDH1A1 genes, nominally significant score differences were observed at select time points in subscale cluster 1 (at later time points) and subscale cluster 2 (peak and just after peak time points) but these did not reach the multiple-testing-adjusted significance threshold (eFigs. 9–11). Of note, the ANCOVA results with the covariates as described in the Methods section did not differ significantly (data not shown).

Worth noting, rs1260326 in the GCKR gene has been associated with alcohol consumption [8, 27]. Our results suggest that this gene has influence on alcohol metabolism itself to some extent, possibly resulting in its association with alcohol consumption. Regarding ALDH1A1 and ALDH1B1, our results align well with their known functions. The affinity to acetaldehyde is ordered as follows: ALDH2 (the strongest), followed by ALDH1B1 and then ALDH1A1. This is consistent with our findings where the ALDH2 variant showed the strongest associations, the ALDH1B1 variant had more significant associations than ALDH1A1 [4].

Thus far, our results suggest that SRs might be significantly shaped by five alcohol-related loci, in order of influence: ALDH2, ADH1B, ALDH1B1, ALDH1A1, and GCKR. This could be supported by the relationship with the PCs, especially PC1, which captures the largest fraction of variance, and the alcohol-related variants, as demonstrated in eTable 14. The strongest associations between participants’ PC1 and subscales in cluster 1 (eTable 15) further support the significant genetic contribution of the ALDH2 variant, given that ALDH2 demonstrated the strongest association within cluster 1. Additionally, to evaluate genotype-related effects on alcohol response at the subscale cluster level, we calculated z-scores from the p-values reported in eTables 6, 7 by comparing the effects of ALDH2 and ADH1B alleles at the 30- and 150-min time points for each subscale cluster (eTable 16). At the 30-min time point, subscale cluster 1 showed the strongest genotype-related differences, followed by subscale cluster 2 and then subscale cluster 3, for both ALDH2 and ADH1B. At 150 min, the ALDH2-related difference in subscale cluster 1 slightly declined, whereas that for ADH1B increased, and both remained stronger than those in subscale clusters 2 and 3. In summary, the subscale clusters consistently exhibited distinct patterns of genotype-related differences across both genes and time points, with subscale cluster 1 emerging as the most genetically responsive domain of alcohol response. These findings suggest that participant cluster 1, which exhibited the strongest alcohol response, may also be the most influenced by genetic variation in ALDH2 and ADH1B, followed by participant clusters 2 and 3.

Based on the results so far and considering the enzymes involved in alcohol metabolism, it is likely that higher acetaldehyde concentrations have a more pronounced impact on SR scores [4]. Notably, the genetic influences appear to segregate into three distinct clusters. Consequently, we propose that the varying contributions of genetic factors, particularly those of ALDH2 and ADH1B, account for the phenotypic differences observed across the three-cluster classifications.

Of note, the assessment of acute adaptation also supports the three-cluster results of the subscales. That is, the susceptibility to acute adaptation differed across the clusters. Without considering the genetic variants, overall, the highest associations with acute adaptations are observed in subscale cluster 2, followed by subscale cluster 3 and then subscale cluster 1 (eTable 17). Furthermore, the evaluation of these associations across genetic variants revealed that subscale cluster 2 lacks any contribution of ALDH2 to acute adaptation (eTable 18). Among participants with T/T vs. T/C or C/C genotypes of rs1260326 in the GCKR gene, T/T vs. T/A genotypes of rs8187929 in the ALDH1A1 gene, and G/G vs. G/C or C/C genotypes of rs3043 in the ALDH1B1 gene, only a few subscales in subscale cluster 1 showed minimal associations with acute adaptation.

Variance component analyses of five loci related to alcohol response

We next performed variance component analyses, which also revealed distinct patterns across the three subscale clusters. Overall, the variance explained by rs671 in ALDH2 peaked at the 30-min point for most subscales in cluster 1, reaching a maximum of 30% for BSS (warmth) in subscale cluster 1 (Fig. 4a, b, eFig. 12, eTable 19). Subscale cluster 1 had a larger explained variance compared to subscale clusters 2 and 3 (Figs. 4a, b). In comparison, the variance explained by ADH1B was typically highest at the 150-min point, especially in subscale clusters 1 and 2, with a maximum of 1.7% for BSS (warmth) in subscale cluster 1 (Figs. 4c, d, eFig. 12, eTable 19). The variances explained by GCKR, ALDH1A1 and ALDH1B1 were generally lower than those of ALDH2 and ADH1B, and thus these three contribute less than ALDH2 and ADH1B (eFig. 13, eTable 19).

Fig. 4. Time courses of attribution of tag SNPs in ALDH2 and ADH1B to each subscale.

a ALDH2 Time courses of the percentage of variance explained by tag SNP in ALDH2 for each subscale in the three clusters. The x-axis represents time points after the start of alcohol infusion, including the period after discontinuation. The y-axis represents the variance explained. Lines in the graph are color-coded for each subscale, with the solid line representing cluster 1, the dotted line representing cluster 2, and the dashed line representing cluster 3, all based on PCA results from the 30-min mark. b Comparison of the attribution of the tag SNP in ALDH2 between the 30-min and 150-min marks after alcohol administration. The x-axis represents subscale, while the y-axis displays the variance explained. c ADH1B. Time courses of the percentage of variance explained by tag SNP in ADH1B for each subscale in the three clusters. The x-axis represents time points after the start of alcohol infusion, including the period after discontinuation. The y-axis represents the variance explained. Lines in the graph are color-coded for each subscale, with the solid line representing cluster 1, the dotted line representing cluster 2, and the dashed line representing cluster 3, all based on PCA results from the 30-min mark. d Comparison of the attribution of the tag SNP in ADH1B between the 30-min and 150-min marks after alcohol administration. The x-axis represents subscale, while the y-axis displays the variance explained.

ALDH2 seems to predominantly influence early alcohol metabolism, while effects of ADH1B manifest later, (Fig. 4 and eFig. 12), contrary to known enzymatic mechanisms (eFig. 1). The contributions of the five loci appear to be time-dependent (Fig. 4 and eFig. 13). Variance partitioning analysis confirmed these genetic contributions, with minor contributions from covariates such as BMI and sex. (eFig. 14). Thus, after alcohol administration, influence of ALDH2 in SR appears most pronounced in the early phases, diminishing over time as the influence of ADH1B be more dominant (eTable 19). No gene exhibited a strong contribution at any time point to the BAES (stimulant) (eTable 20).

Polygenic architecture of SR

We further investigated unreported genetic associations and potential polygenic architectures of SR. We performed GWAS for each subscale, taking into account the timing of peak scores in each subscale (eFig. 3). ALDH2, which contains rs671, was the only GWAS-significant region without inflation of association statistics (data not shown). Next, we constructed PRS for each subscale, excluding regions of rs671 in ALDH2, rs1229984 in ADH1B, rs1260326 in GCKR, rs3043 in ALDH1B1, and rs8187929 in ALDH1A1 as well as the regions in linkage disequilibrium with those SNPs. This was done to gauge the polygenic effects of unknown regions and avoid overfitting (See Methods). There were no significant correlations between scores across any subscales after cross-validation. While the current dataset is underpowered, this may suggest a limited polygenic burden outside the known alcohol-related genes on SR.

Discussion

This study is the first to demonstrate that responses to alcohol intake and individuals can be classified into three distinct participant clusters, each of which exhibits a unique profile of alcohol response. Each individual cluster is characterized by genetic contributions related to alcohol metabolism and consumption, with these contributions varying in timing and extent across clusters. This was achieved by comprehensively examining the relationship between pharmacodynamics (specifically, SR) and genetic loci related to alcohol metabolism, and by measuring three indicators of alcohol reactivity at consecutive time points within a large, homogeneous group of healthy individuals. Additionally, the inclusion of young participants is another advantage, as younger participants typically have stable phenotypes with minimal influence from non-genetic factors [28, 29].

Our study also offers new findings different from previous research [30, 31]. These prior genetic studies utilized SR as a phenotype, demonstrating that genetic variation in alleles for alcohol-metabolizing enzymes is associated with differential SR. For instance, Marshall et al. [30] investigated the effect of ADH1B polymorphisms on SR in African Americans. Similarly, Chen et al. [31] examined East Asian populations, highlighting the associations and variations in SR based on ALDH2 polymorphisms. Furthermore, a study conducted by one of our co-authors demonstrated associations between GABRA2 polymorphisms and SR [21]. In contrast, our study focused on alleles prevalent in the Japanese population [3], aiming to explore the precise contributions of genes to SR as a consequence of differences in alcohol metabolism and consumption, which remains an unexplored area in prior research. According to a previous review, the proportion of variance explained ranges from 4 to 13% [32]. In contrast, our study shows a 30% variance explained in BSS (warmth) by rs671, within Japanese populations, due to the East Asian-specific nature of rs671 (refer to eTable 21). We also identified the importance of another East Asian-specific locus, rs8187929 in ALDH1A1 given substantial variance explained depending on the time point after alcohol intake. Additionally, our study was the first to examine the polygenic architecture of SR, and it demonstrated a limited polygenic burden outside the known alcohol-related genes regarding alcohol responsiveness.

Our findings challenge the assumed sequential activation of ADH1B followed by ALDH2 in alcohol metabolism. Contrary to expectations, ALDH2 may act earlier in this process. This could be due to multiple roles of ALDH2 beyond its established metabolic functions, potentially altering early alcohol processing [33]. Additionally, SR may not directly reflect alcohol and acetaldehyde levels or their relationship may not be linear, possibly influenced by more efficient enzymatic efficiency of ALDH2, which lowers Michaelis constant (Km) for acetaldehyde compared to that of ADH1B [34]. These findings needs to be validated through detailed functional studies.

The composite measures employed in our study to assess SR have been previously well-studied for their clinical implications. For instance, BAES has been associated with the prognosis of alcohol-related disorders [35]. However, our findings revealed very weak contributions of the genes investigated in this study to the BAES stimulant response. Additionally, variance partitioning analysis indicated minimal contributions from age, sex, and BMI. This lack of strong genetic or demographic association suggests that other, as-yet unidentified factors may play a critical role in shaping the stimulant response. Similarly, a low SHAS score has been linked to an increased likelihood of developing alcohol use disorder [36]. In this study, we found that these scaling measures can also be divided into three distinct clusters. Additionally, our assessment of acute adaptation—including analyses across genetic variants—supports these three-cluster classifications. By using this classification, we could better identify groups at risk for alcohol-related disorders.

Our study has limitations. While this is the most extensive examination of the relationship between genetic architecture and SR using three assessment tools, the sample size remains a constraint. In this study, we included individuals with a history of drinking, which led to a lower frequency of the ALDH2*2/*2 genotype compared to the general population; however, this limitation does not compromise the generalizability of our findings, as the study focuses on the response to alcohol, which is inherently inapplicable to those unable to consume alcohol. Additionally, despite our comprehensive assessment of the polygenic architecture, incorporating more rare variants through next-generation sequencing in a larger sample could improve predictive performance [37]. Despite these considerations, our work offers valuable insights by demonstrating that the three-cluster classification, which captures variations in the degree and timing of alcohol response among carriers of the same allele, could inform future strategies for identifying individuals at risk for alcohol-related disorders. While this approach is not yet ready for direct clinical application, these findings enhance our understanding of the genetic factors influencing alcohol responsiveness, thereby laying the groundwork for potential future interventions.

Author contributions

KH, SM, and CT made substantial contributions to the conception and design of the study. SM contributed to the acquisition of data. KH, SO, SM, and CT contributed to the analysis of data. KH, IO, SO, AH, KO, XL, YI, TM, SM, and CT contributed to the interpretation of data. KH drafted the manuscript, and IO, SO, AH, KO, XL, YI, TM, SM, and CT critically revised it for important intellectual content and approved the final version for publication. All authors agree to be accountable for all aspects of the work, ensuring that any questions related to accuracy or integrity are appropriately investigated and resolved.

Funding

This work was supported by Asahi Quality & Innovations, Ltd., Japan Agency for Medical Research and Development (AMED) grants 21ek0109555, 21tm0424220, 21ck0106642, 23ek0410114, and 23tm0424225, Japan Society for the Promotion of Science (JSPS) KAKENHI grant JP20H00462.

Data availability

Summary Statistics will be available in JENGER (http://jenger.riken.jp/en/).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41386-025-02147-7.