Lifestyle and environmental exposures have been shown to modify the associations of common genetic variants with traits linked to cardiometabolic disease (Grarup et al., 2008; Montasser et al., 2009; Higashibata et al., 2012; Manning et al., 2012; Sung et al., 2014). Recent large-scale studies have expanded the genetic architecture of interaction effects with genome-wide interaction analyses with tens of thousands of people from diverse ancestral backgrounds (Feitosa et al., 2018; Sung et al., 2018; Bentley et al., 2019; de Vries et al., 2019; Sung et al., 2019). These studies implicate genetic loci for which the genetic effect on a phenotype of interest is modified by either alcohol consumption or cigarette smoking. These lifestyle exposures could impact the gene regulatory mechanisms through which the trait-associated alleles influence gene expression. Statistical tests that evaluate these models would require human cohorts with exposure and outcome data, genetic data, and methylation, gene expression or other ‘omics data. The process of finding, obtaining, harmonizing, and analyzing individual-level data from human cohorts presents a challenge, as these data are ‘controlled-access’ and require numerous regulatory steps for the researcher, the researcher’s institution and the entity managing access to the data. On the other hand, it is now common to publish summary-level association statistics from human cohorts. Therefore, an analysis of both genetic and non-genetic data that relies on summary data alone is particularly valuable in this context.

To this end, we leveraged multiple summary-level association datasets, both genetic and epigenomic, to evaluate possible mechanisms of gene-environment interactions effects. We use data from the CHARGE Consortium’s Gene-Lifestyle Interactions Working Group (Rao et al., 2017), published in five recent papers (Feitosa et al., 2018; Sung et al., 2018; Bentley et al., 2019; de Vries et al., 2019; Sung et al., 2019) together with differential methylation and gene expression data (Lonsdale et al., 2013; Dekkers et al., 2016; Huan et al., 2016; Joehanes et al., 2016; Parker et al., 2017; Richard et al., 2017; Liu et al., 2018; Nikodemova et al., 2018). We demonstrate that by considering diverse epigenetic associations at the same genetic loci and linking multiple forms of evidence to a common gene, gene-environment interaction effects may be more fully characterized and prioritized for future follow-up analyses.

Although gene × environment interactions are often cited as a potential source for “missing heritability” (Simon et al., 2016; Mayhew and Meyre, 2017), molecular evidence of these interactions remains limited. The establishment of CHARGE’s Gene Lifestyle Interactions Working Group resulted in a number of large-scale, trans-ancestry evaluations of gene × environment interaction effects (Feitosa et al., 2018; Sung et al., 2018; Bentley et al., 2019; de Vries et al., 2019; Sung et al., 2019). In this analysis, we leveraged publicly available epigenomic and transcriptomic summary association data to identify a subset of interactions from these analyses among which multiple forms of evidence point toward potential mechanistic explanations. We identified 833 genes for which evidence of a statistical interaction can be supported with functional data from at least one source, with 48 of these having multiple sources of functional support. We also observed an overrepresentation of prioritized interaction loci drawn from meta-analyses of African-ancestry populations. An overrepresentation of findings with the smoking vs. alcohol exposure may be a result of the greater amount of available omics data for smoking (Table 1).

A striking finding from this work is the large proportion of interactions that are observed among African ancestry meta-analyses compared to other ancestries or to Trans-ancestry analyses. This phenomenon was observed among both the full number of interactions identified (79.7%) as well as the prioritized loci (62.5%), and three of the five further prioritized loci were from African ancestry meta-analyses. This phenomenon reflects the underlying GLI association results from which our study draws. Notably, despite the shared methodology, this predominance of findings from African ancestry meta-analyses was evenly distributed among traits considered but was not evenly distributed among exposures, observed in studies of smoking with blood pressure and lipids, but not with alcohol exposure on either trait. Similarly, in these analyses, 26 of 30 of the prioritized loci based on meta-analyses of those of African ancestry were studies of the smoking exposure. A more complete discussion of these GLI association results are available in the primary publications for these projects. Briefly, some of the associations are for variants that are in higher frequency or present only among African ancestry populations, they are generally of low frequency (MAF 0.01–0.05) with high imputation scores, and with consistent associations across contributing African ancestry cohorts (Sung et al., 2018; Bentley et al., 2019; Sung et al., 2019; Laville et al., 2020). Of the 30 loci prioritized in this study based on African ancestry meta-analyses, the lead associated variant was African ancestry-specific (only present in 1 KG AFR populations) for only 2, while for most the lead variants were available in all ancestries, but not associated in the meta-analyses of other ancestries. These results suggest that the source of this observation relates to a smoking exposure-related difference by ancestry.

Although we did not have the data for a detailed evaluation of smoking patterns by ancestry in our studies, there are pronounced differences in smoking patterns across ancestry groups in the US (U.S. Department of Health and Human Services, 1998). Notably, there is a marked difference in the type of preferred cigarette, as shown in data from the National Survey on Drug Use and Health, which is designed to be representative of the US population: 88% of African Americans smokers vs. 26% of non-Hispanic white smokers used menthol cigarettes (Giovino et al., 2015). Menthol cigarettes are marketed more aggressively to African Americans (Lee et al., 2015; Mendez and Le, 2021). Additionally, some differences in preference may stem from variations in bitter taste perception by ancestry, with menthol cigarettes more palatable to those with stronger bitter taste perception as the menthol flavor additive masks the bitterness of nicotine. An African ancestry-specific genetic locus, MRGPRX4, associated with a five- to 8-fold increased odds of menthol cigarette smoking was recently identified (Kozlitina et al., 2019), although only a small minority of African Americans carry this variant.

Menthol cigarettes have long been targeted by the public health community based on evidence that they facilitate deeper smoke inhalation by decreasing nicotine-induced irritation (Ton et al., 2015; Mayhew and Meyre, 2017). This deeper inhalation may lead to a subsequent higher absorption of the myriad harmful components within cigarette smoke (U.S. Department of Health and Human Services, 1998; Ross et al., 2016). Consistent with this observation, ancestry differences in smoking-related metabolites and carcinogens have been reported (Pérez-Stable et al., 1998; Benowitz et al., 2011; Khariwala et al., 2014; Jain, 2015), and different levels of these compounds may underlie the observed differences by ancestry in genetic interactions upon smoking exposure. Additionally, there is some evidence for greater systemic oxidative stress (Il’yasova et al., 2012; Morris et al., 2012; Annor et al., 2017; Kim et al., 2020) and inflammation (Albert et al., 2004; Khera et al., 2005; Akinyemiju et al., 2019; Zahodne et al., 2019) among Americans of African vs. European ancestry. Exposure to cigarette smoke, a rich source of oxidants, on a background of elevated oxidative stress and inflammation may provoke a greater response among these individuals, manifesting as an interaction with smoking that differs by ancestry.

One key motivation for conducting analyses of gene × lifestyle interactions is the relative ease of practical translation, as results suggest a modifiable risk, i.e. individuals with a certain genotype might reduce exposures associated with exacerbated risk. In our efforts to map existing functional information to loci of interest, we identified several areas where improvements in available data might facilitate stronger inferences. For instance, there are limited data to evaluate differential gene expression by alcohol exposure, making it difficult to further investigate the loci identified in gene-alcohol interaction analyses. Additionally, more tissue-specific data would be useful. Specific tissue types are of greater interest for each phenotype (e.g. liver for lipids) and for each exposure (e.g. lung for smoking). Further, the reliance on whole blood, which is the most available in cohort studies, will limit our understanding of the underlying biology. Data linking methylation to gene expression is also limited. Although it was beyond the study design of this project, it would be useful to collect individual-level data to better elucidate these loci. Importantly, while some of the patterns observed in our data fit with expectations (e.g. the direction of RNA expression and DNA methylation for GCNT4, PTPRZ1, TMEM116), some did not, and individual-level data is needed to correctly model these associations.

A strength of these analyses is the use of the expansive genome-wide interaction results from large epidemiological meta-analyses. These data are drawn from discovery data on up to 133,805 individuals, important given the statistical power needed to detect gene-environment interaction effects. Additionally, the CHARGE Gene × Lifestyle Interactions Working Group went to great effort to include studies of diverse ancestries, such that relatively large proportions of historically underrepresented ancestry groups, such as African ancestry, were achieved. Given the preponderance of African ancestry-identified associations among our results, this inclusion was of key significance. This work could have been improved with omics data derived from individuals of diverse ancestries to explore associations that differed by ancestry. The exposure data considered in these analyses was represented using binary variables in order to maximize sample sizes for detection of interactions, although the true effects of exposures are certain to be more complex, with variations with timing and dose of exposure. Similarly, the exposures and phenotypes we selected for these analyses were relatively straightforward to measure, however, a wide range of exposures may be involved in gene-lifestyle interactions on a wide range of phenotypes, and it is unknown whether these findings are representative of gene-lifestyle interactions in general. Also, additional experimental data will be necessary to confirm biological pathways suggested by these findings in order to advance the evidence from this work towards clinical translation.

In summary, this work provides preliminary evidence from publicly available transcriptomics and epigenomics data to support the biological mechanisms underlying statistical gene × lifestyle interactions. These data suggest compelling evidence for how gene × lifestyle interactions may occur, motivating future studies that include individual-level epigenomic and transcriptomic data, other environmental exposures and outcomes, and more complex characterization of exposure.