While the field of AUD genetics has made considerable progress, substantial gaps persist (similar to other psychiatric disorders, ref. 91). Here, we highlight some limitations of the current AUD studies (Table 2) with the hope that gaps may be filled with new data sets, technologies, analytic methods, and research directions in the future.

Challenges of genetic studies of AUD

(a) Different definitions of AUD and proxy phenotypes (e.g., AUDIT-P) have shared genetic architecture, resulting in improved power in gene discovery when they are combined from different cohorts (78, 80). However, they are not identical traits. Deep phenotyping (either using same definition or focusing on subphenotypes) in larger cohorts could reduce the phenotypic heterogeneity and increase the possibility of identifying trait-specific associations and pathways (92).

(b) Some studies have endeavored to include samples in multiple ancestries (55, 56, 63, 64, 75, 80), but the sample sizes in the non-European ancestries are smaller than sample sizes in the European ancestries — a common issue in human genetic studies (53, 54). Recruitment of individuals of diverse genetic ancestries is a critical next step in this field. With more multiancestral biobanks becoming available, including MVP, the Global Biobank Meta-analysis Initiative (93), and the All of Us Research Program (94), we anticipate that the gap in diversity will diminish. Funding agencies should also direct attention to studies that propose recruitment focused on non-European ancestry participants.

(c) AUD is a highly polygenic disorder, with hundreds of variants at least contributing to the risk (80, 95). The “brute force” GWAS approach requires a larger sample size to identify more risk variants. Unlike other traits or behaviors that can be measured directly and assessed in large populations or biobanks — for example, GWAS of height (96), educational attainment (97), and alcohol consumption (98) have been conducted in 3~5 million participants — clinical diagnosis of AUD in large cohorts is still lagging. Similar to point (a), increasing sample size and incorporating multiple ancestries could improve the power and resolution of causal variant fine-mapping (80). Besides the well-known functional coding variants in the alcohol metabolic genes, most variants identified through large GWAS have small to very small effects on the risk of AUD, reducing the yield of the extensive effort of following functional studies on individual variants. This is a common issue in the genetic study of complex traits.

(d) Current GWAS studies have mostly used SNP arrays and post hoc imputation to fill in common variants, which does not allow analysis of the full genome because some parts of the genome are not fully “covered” — i.e., there are unassessed variants in some genomic regions that cannot be tested for association, for technical reasons. A typical SNP array can capture from 600,000 (for example, Illumina PsychArray) to 1.8 million (for example, Illumina Multi-Ethnic Genotyping Array) variants. After imputation and application of standard quality controls for the variants, typical analyzable numbers of high-quality variants vary from 5 to 15 million, depending on the original array SNP density, sample size, and genetic ancestry (from a population genetics point of view, African populations have more common variants than other populations due to their evolutionary history). Given the inherent missing information from different steps, GWAS meta-analyses can only cover a subset of variants of the whole genome, indicating that much of the genome is missing in the current genetic studies of AUD. Whole-genome sequencing (WGS), which can detect essentially all variants (including rare variants and structural variants) without ascertainment bias, could provide better opportunities to investigate the full genetic architecture of the trait.

Several whole-exome sequencing (WES) studies and one WGS study of AUD have been conducted recently (99–102). A phenome-wide WES study of 170,979 individuals (6,320 cases) from the UK Biobank identified two common variants in the ADH1C gene associated with AUD (P < 2 × 10–9), using either an additive or a dominant model (102). A WES study combining 469,835 individuals from the UK Biobank data (13,121 cases) and 3,789 individuals from the Yale-Penn cohort (2,562 cases) with multiple ancestries identified the well-known functional variant ADH1B*rs1229984 and several common variants in ADH1C. Gene-based tests accounting for the burden from loss-of-function, missense, and synonymous variants identified novel genes CNST and IFIT5 (101). A low-coverage WGS study of AUD-related life events and two affective symptoms in 742 American Indians and 1,711 European Americans identified both common and rare novel variants (103).

(e) Most variants identified by GWAS are in noncoding regions with unknown functions (104). The top associated variants in each risk locus are not necessarily the causal variants for AUD. Although post-GWAS fine-mapping analysis could identify a credible set of potential causal variants (105–107), further efforts are needed to interpret and validate the variants’ functions. In recent years, novel analytic approaches like deep learning (a subset of machine learning) have been successfully implemented in biomedical research. For example, deep-learning methods contribute to prediction of protein structure (108, 109), pathogenic missense variants (110, 111), and regulatory functions of genome variations (112–115). Combining novel computational tools and cutting-edge functional essays like genome editing (116–118) could help assess the variants’ effects at scale.

(f) Although hundreds of risk variants have been identified and many have been repeatedly replicated in GWAS, indirect genetic effects (also called “genetic nurture”), which are effects of alleles in parents on offspring through the environment (119), have not been distinguished from direct genetic effects on AUD. Methods have been developed to impute parental genotypes using family data (120), which could be used to improve estimates of direct genetic effects for AUD. Confounding effects, including socioeconomic status, may also bias the results. For example, educational attainment influences many psychiatric and nonpsychiatric traits (97) and has a genetic correlation rg = –0.21 with AUD, which needs to be considered in future studies.

(g) Another profound gap is that the current predictive performance of PRS for AUD based on GWAS common variants — i.e., using genetic variation to predict risk in genotyped individuals — is strongly statistically significant but numerically still weak and has not yet entered the range of clinical utility. Despite the increase in sample size, the SNP-based heritability (h2) by GWAS is low (h2 ranges from 5.6% to 12.7% with liability-scale h2 ranging from 8.9% to 16.2%, refs. 64, 72, 75, 78, 80) compared with the total heritability but comparable to what is observed for many other genetically complex traits. PRS presently has limited power for AUD prediction (explained variance measured by pseudo R2) in independent cohorts; thus, the clinical use of the current PRS of AUD is not imminent. Possibly, the success of artificial intelligence in other areas could extend into predicting AUD risk, with more genomic and large-scale electronic health records data available by integrating improving genomic data with other trait predictors.

(h) Finally, genetic studies have confirmed that AUD is partly a brain-related disorder (75). Genes with expression perturbation in specific brain tissues have been prioritized (72, 80), but the biological pathways from genetics to the etiology of AUD are largely unclear. There are major exceptions though: the mechanism of the effect on risk of alcohol-metabolizing enzyme variation is well understood. Many biological processes play roles in the pathways, such as gene expression, functional regulation, protein perturbation, metabolites, and other mediating traits. To understand the pathway mechanisms, studies beyond genetics are warranted, including, but not limited to epigenetics (discussed in Epigenetics of AUD), multiomics, single-cell sequencing, and the latest spatial transcriptomics.