Current Challenges and New Opportunities for Gene-Environment Interaction Studies of Complex Diseases

Abstract

Recently, many new approaches, study designs, and statistical and analytical methods have emerged for studying gene-environment interactions (G×Es) in large-scale studies of human populations. There are opportunities in this field, particularly with respect to the incorporation of -omics and next-generation sequencing data and continual improvement in measures of environmental exposures implicated in complex disease outcomes. In a workshop called “Current Challenges and New Opportunities for Gene-Environment Interaction Studies of Complex Diseases,” held October 17–18, 2014, by the National Institute of Environmental Health Sciences and the National Cancer Institute in conjunction with the annual American Society of Human Genetics meeting, participants explored new approaches and tools that have been developed in recent years for G×E discovery. This paper highlights current and critical issues and themes in G×E research that need additional consideration, including the improved data analytical methods, environmental exposure assessment, and incorporation of functional data and annotations.

Genetic and environmental factors are thought to contribute to the etiology of most complex diseases. Through genome-wide association studies (GWAS), thousands of common loci associated with complex diseases have been identified (1–3). Researchers have been motivated to discover and describe how the interplay of these factors influences disease risk and outcomes. There are several reasons to study gene-environment interaction (G×E): providing insights into the biology of disease (e.g., developing new models for disease etiology based on observed G×E findings), building better prognostic models (e.g., using genotype to inform treatment and prognosis), identifying possible high-penetrance subgroups (e.g., increased genotype-specific risk in premenopausal women), or identifying genetic subgroups with higher exposure-specific disease risk for prevention efforts (e.g., increased environmental-specific risk for individuals with a particular genotype) (4–7). Furthermore, in the search for novel genes via GWAS, the modifying effects of environmental risk factors are not often taken into account; therefore, leveraging G×E may result in discovery of additional disease susceptibility loci (5, 8, 9). Despite interest in G×E, there are few agreed-upon successes where the effect of exposure differs across genotypes (and vice versa). Numerous reasons have been suggested as contributors to the small number of successes, including the inherently low power of tests for G×E, the complexity of measuring environmental exposures, the difficulty of incorporating the temporality of environmental exposures, measurement error, limited range of genetic and/or environmental variation, scale dependence in the definition of statistical interaction, and lack of data on the biological consequences of most genetic variants (10–13).

The past few years have seen an emergence of new approaches, study designs, and statistical and analytical methods for exploring G×Es in large-scale studies of human populations. Further, new opportunities in this field continue to be developed with respect to the incorporation of -omics and next-generation sequencing data and improvements in measures of environmental exposures implicated in complex disease outcomes. Therefore, on October 17–18, 2014, the National Institute of Environmental Health Sciences and National Cancer Institute held a workshop at the 64th Annual Meeting of the American Society of Human Genetics—“Current Challenges and New Opportunities for Gene-Environment Interaction Studies of Complex Diseases”—to explore these new approaches and tools for G×E discovery. Based on the discussions, we prepared 4 articles that provide an update on: 1) the state of the science in analytical methods (14); 2) opportunities for incorporation of biological knowledge into G×E analyses (15); 3) advances in environmental exposure assessment in human populations (16); and 4) lessons learned from G×E successes (17). In addition, this article develops some overarching themes and sets the stage for this series. Because environmental factors may be modifiable, defining subpopulations of individuals that are most susceptible to environmental factors through G×E analysis may provide targets to improve public health. This idea is consistent with the goal for President Obama's recently launched Precision Medicine Initiative at the National Institutes of Health: to better understand how individual variability contributes to differences in response to treatment or prevention (18, 19).

ANALYTICAL METHODS

G×E studies require much larger sample sizes than studies targeting either genetic or environmental main effects alone (20). Further, when performing G×E studies on a genome-wide scale, sometimes referred to as genome-wide interaction studies, sample size requirements are substantially further inflated to account for the multiple comparisons performed (5, 21). Therefore, a goal of the development of G×E methods has been to improve power to detect associations. As detailed in Gauderman et al. (14), many different methods have been explored in the context of case-control studies as alternatives to traditional G×E tests, including case-only studies (22), empirical Bayes (23), Bayesian model averaging (24), joint tests (9, 25, 26), case-parent approaches (27–29), and 2-step approaches (21, 25, 30–35). Other approaches include set-based methods, which combine multiple variants or G×Es and which may be particularly appropriate for studies of rare variants (36–40). In addition, several methods have been developed to analyze G×E for quantitative outcomes (41–48).

The large number of available methods, as well as novel software tools to support the application of these methods (31, 49, 50), creates opportunities to better study G×Es in genome-wide settings. Researchers may therefore wonder which method to use for their studies. Several previous simulation studies suggest that none of these G×E methods is universally the most powerful approach (31, 32, 51–54). Therefore, decisions about the most appropriate approach depend on several considerations, including the hypotheses to be tested, likely genetic architecture, study design attributes, and characteristics of the population being studied. Investigators should be cautious about applying multiple methods to their data without an a priori basis for choosing among the results—simply picking those with the most “significant” findings to report would clearly be a biased strategy that could contribute to spurious associations and to what has been referred to as a “vibration of effects” (55, 56). Some of the new methods, however, provide flexible frameworks for combining multiple tests with an appropriate permutation procedure to evaluate the significance of the overall results (31, 32). The collection of methods allows investigators to address specific scientific questions and offers new opportunities for studies of G×E in large populations.

FUNCTIONAL VALIDATION AND DISCOVERY

Despite the recent success of GWAS at identifying risk loci, variants identified are by design not usually the causal variants, defined as the functional genetic variant that influences risk of disease and explains the association. Currently, the underlying biological mechanism contributing to disease risk is only known for a small proportion of these loci. Therefore, more research to functionally characterize risk loci is now being performed, providing opportunities by which G×E analyses may offer new insights into disease development (57). An understanding of the biological consequences of particular genetic differences could lead to specific mechanistic hypotheses, identifying relevant exposures to test and specifying relevant statistical models. As described in Ritchie et al. (15), these approaches include using functional annotations for discovery and validation, studying molecular phenotypes (e.g., epigenetics or gene expression) to improve G×E discovery, and leveraging in vitro and in vivo models for these studies.

Several large public databases (such as the Encyclopedia of DNA Elements (ENCODE), Epigenomics Roadmap, Genotype-Tissue Expression (GTEx), and the Cancer Genome Atlas (TCGA)) have facilitated the functional annotation and interpretation of many genomic regions, which can be used to prioritize candidate G×E markers (32). Many disease-associated single-nucleotide polymorphisms identified in GWASs appear to be located in noncoding or regulatory regions, which are often affected by environmental exposures (58–60). The Encyclopedia of DNA Elements and Roadmap Epigenomics programs have helped to define many of the regulatory regions, and new tools developed by these programs and others now allow functional annotation information—such as the genomic location of histone modification states, methylation patterns, transcription-factor binding sites, and DNAse hypersensitivity sites or other higher-order chromosomal structural information—to be overlaid with GWAS results and could be integrated into G×E analyses (61–65). Projects like Genotype-Tissue Expression have greatly increased the compendium of putative biological functions of genetic variants. However, neither Genotype-Tissue Expression nor large-scale epigenomics projects provide information on effects of genetic and genomic functions across a range of environmental conditions. To explore genetic effects in response to environment, in vitro studies have now perturbed cells and recorded responses to various drugs, infections, and other exposures. Through the use of intermediate molecular phenotypes such as gene expression, these efforts have demonstrated success in illustrating how an exposure may influence gene function, suggesting potential candidate genes or variants for G×E studies (66–70).

In addition to data resources, the use of population-based mouse resources (such as the Collaborative Cross, Diversity Outbred, and Hybrid Mouse Diversity Panel) and other appropriate mouse models have also been leveraged to assist in the discovery or replication of G×Es. These population-based, variant-enriched mouse resources have been designed to mimic the genetic diversity of human populations and can be used to replicate or inform G×E hypotheses by using carefully controlled exposures in mouse studies. Several recent examples have exemplified the power of these resources to map genetic variants related to susceptibility to environmental exposures (71, 72). Although both in vitro and model systems have led to potential mechanistic insights, linking of these to human populations remains challenging.

There are many approaches for incorporating biological knowledge to improve analytical methods (73) for G×E in both the discovery and the validation phase. Incorporating functional annotation data and a priori biological information (such as knowledge on biological pathways and metabolomics or gene-expression data collected on individuals) to inform methods for analyzing G×E data has aided in the discovery of new G×E findings in recent years (74). For example, Bayesian variable selection (75, 76), the Algorithm for Learning Pathway Structure (77), and the PEAK algorithm (74) are all methods that incorporate external biological information and properties of the data set itself to increase power over agnostic approaches to detect interactions. Another approach is to use 2-stage modeling where functional annotations are used to prioritize variants (78, 79) for G×E studies. As one example, Biofilter was designed to build biologically plausible models of gene-gene and G×Es to test for associations based on biological features using biological knowledge from the public domain (78, 80). These types of filtering approaches are also being explored to prioritize environmental exposures by using databases such as the Comparative Toxicogenomics Database, which links exposures to genes (81). However, challenges still exist in linking environmental exposures into currently available ontological knowledge resources, although some investigators are beginning to navigate these challenges (82). Furthermore, all these databases and functional annotations depend on the quality and extent of existing biological knowledge (73).

ENVIRONMENTAL EXPOSURES

The complex realities of environmental exposures have long made measurement of exposures substantially more complicated than inherited genetic measurements (e.g., genotypes) and single nucleotide variants in particular. The technologies and approaches to incorporate exposures into human population studies have therefore lagged behind genomics capabilities (11, 83). Assessing exposure impact must take into context not just the variety of exposures themselves (physical, often complex chemical mixtures, biological, and psychosocial) but also the source and place of exposure, the timing during a person's life trajectory, the route of contact (skin, lung, diet), metabolism/excretion, and distribution in target tissues. All of these factors may affect the ultimate disease risk associated with environmental exposures. In addition, in the classic environmental exposure paradigm, studies may focus on measurements to capture internal versus external exposure, early markers of disease, or an ultimate biological response, which further adds to the complexity of exploring the impact of environmental exposures.

In recent years, however, exciting new opportunities have become available for environmental exposure assessment. The potential importance of examining the totality of internal and external exposures, referred to as the “exposome,” has been recognized (83, 84). Several recent commentaries have described considerations for measurements of the exposome (85–88). Innovative technologies—including activity monitors, improved sensors, global positioning systems, and geographic information systems—enable new and more detailed exposure measurements, although issues of the timing of exposure measures persist and should be considered. Moreover, development of biological response markers for assessment of exposure (such as changes in gene expression, transcriptomic signatures, and DNA methylation profiles) has been useful for G×E discovery (89–92). Another opportunity is the exploration of environmental exposures in a more agnostic discovery-based fashion, similar to GWASs. These studies, termed environment-wide association studies (EWASs), have led to discoveries of environmental factors associated with disease (93–96).

Key challenges and considerations remain associated with assessing environmental exposures in G×E studies (16), including how to select the most appropriate study designs, incorporate high throughput -omic measures (e.g., metagenome, metabolome) and sensor technologies into human population-based studies, assess long-term exposure, integrate a variety of divergent external exposure and internal response data, and further advance statistical approaches to handle the dynamic nature of exposure data. We are now at the early stages of exploring which novel exposure assessment technologies can be appropriately applied to larger population studies most effectively. To this end, some 2-stage study designs have been investigated (26, 97–100). Given the extreme cost of incorporating some sophisticated environmental measures into large-scale studies of human populations, the question of what can be accomplished with dense (i.e., repeated measures of a marker or measurement of multiple analytes using an -omic platform) environmental measures on a subsample and extrapolating to a larger sample size—and whether simulations can demonstrate that this approach increases power to detect G×E—is currently being explored (51, 97, 101).

Several analytical methods have been developed for the unique considerations of exposure assessment. New statistical methods can adjust much better for exposure misclassification (which has been shown to lead to inflated type I errors and substantially reduced power). These approaches should allow for obtaining greater power with smaller sample sizes. In addition, novel statistical methods have been developed to detect gene × longitudinal exposure interactions by taking into account long-term time-varying exposures (102). Importantly, as researchers begin to combine exposure data to obtain the larger sample sizes required for G×E research across studies, they have to address the possibility that exposures were measured using different approaches or have very different distributions in and between populations, such that exposure misclassification could produce spurious associations (14). There is also the challenge of exposure-related population stratification for studies relating to G×Es (103). Meanwhile, multiple measures can sometimes increase power for detecting associations. For example, in a recent study, continuous monitoring was shown to reduce the sample size required in a clinical trial context (104).

G×E EXAMPLES FROM HUMAN POPULATION STUDIES

By examining G×E successes, it may be possible to improve the design of G×E studies for the future. Examples of G×E successes range from Mendelian-like traits (e.g., phenylketonuria) to complex diseases (NAT2 variants, smoking, and bladder cancer) and response to therapies (HLA-B*1502 variant and carbamazepine-induced Stevens-Johnson syndrome) (17). In addition, several recent studies examined the use of polygenic risk scores, generated from common genetic variation, to assess the impact of environmental factors on individuals with low versus higher genetic risk (105–108). In Ritz et al. (17), highlighting some of the most successful G×Es identified to date, several common themes have emerged, including the strength of focusing on metabolic pathways for a specific exposure; the utility of studying unique, highly, or diversely exposed populations; the necessity of using high-quality exposure assessment methods; the need for large sample sizes; and the utility of model systems to demonstrate genetic function when replication is challenging in population-based studies. These suggest important avenues for undertaking successful future research in G×E.

THEMES AND FUTURE DIRECTIONS

Inclusion of diverse populations may facilitate G×E research by improving power for discovery of causal genetic variants and environmental factors associated with disease. Transethnic differences in the distribution of linkage disequilibrium can be leveraged to improve fine mapping to identify potential causal alleles (109–112). Combining admixture mapping with conventional GWASs may also facilitate discovery of novel loci (113). Using this latter approach, novel loci were identified associated with total immunoglobulin E levels (114) and asthma (115). Last, using geographically diverse populations might expand the distribution of the environmental exposure and thus increase power to detect interactions (13). Performing genetic studies on populations of diverse ancestry might improve our understanding of disease mechanisms, and such studies are required to ensure all populations benefit equally from this research (116).

Replication is an essential component of genetic association studies, and the requirement for independent replication contributed to the success of GWASs (117, 118). However, replication and meta-analysis become challenging as G×E studies become sophisticated in analytical methods, exposure assessment, and incorporation of functional information. Differences in the underlying distribution of environmental exposures, patterns of linkage disequilibrium, and genetic modifiers can reduce the power to detect the same level of interaction in independent studies. Moreover, an appropriate human replication study might not (yet) exist in studies of a rare disease, genetic variant, or environmental exposure; where exposures are unique to particular populations; or where the initial finding was obtained within a large consortium comprising all known studies of a specific outcome (12). As illustrated in Ritz et al. (17), describing G×E successes and incorporation of biological knowledge, in some situations functional studies could serve to provide support for initial G×E observations in absence of a suitable replication population. Moreover, as the field considers gene- and pathway-based approaches to study G×E, replication may become further complicated as different combinations of genes in different data sets may be observed in the interaction. Some have argued that replication requirements might be met if the underlying biological pathway is the same even if replication was not observed with the individual single-nucleotide polymorphism or gene (15). More consideration is needed of standards for replication, definitions of replication, and alternative approaches for replication and verification of G×E results.

Many exciting opportunities exist for studies of G×E. There is the emerging recognition that developmental exposures may lead to disease throughout life, and efforts have focused on beginning to address how much of the environmental exposure risk for many disease outcomes may be attributable to in utero exposures or other particularly vulnerable windows of susceptibility (childhood, adolescence, etc.). Successful integration of large volumes of diverse data types (including data from geographic information systems, sensors, metabolomics, and other -omics) will create opportunities to generate unique insights. Epigenetics tools open up new opportunities to directly link environmental exposure to the genome and generate new exposure biomarkers (e.g., methylation of cancer-specific genes associated with dietary folate and alcohol in colorectal cancer (119) or smoking exposure in lung cancer (120), as discussed in reviews of opportunities in epigenetic epidemiology (121, 122)). Moreover, epigenomics, as well as other -omic technologies, may elucidate mechanisms by which exposures contribute to disease. The role of the microbiome as a key environmental risk factor for many complex disease phenotypes is starting to be appreciated and extensively studied. In addition, molecular phenotype data creates opportunities to examine disease subtypes or more precisely classify disease. This may eventually reduce heterogeneity in studies and improve the power to study G×E associations, assuming molecular characterizations are performed with the correct cell type, tissue, or appropriate surrogate tissue for the hypothesis being tested.

Additional areas of research may allow further advances in G×E discovery and replication. The field needs to determine how to best leverage experimental studies in animals or human cell lines to aid in discovering and functionally validating G×Es. Moreover, it is unclear how to best leverage existing family and twin-based studies for examining G×E. In incorporating functional information into G×E studies, questions remain about the appropriate balance between using prior or external information and the characteristics of the data set being studied when building analytical models and appropriate methods for linking environmental exposures information into available biological knowledge databases, which are usually focused on genes and pathways. In addition, because many G×E findings to date have modest effect sizes or have not been extensively replicated (11, 123), it is worthwhile to explore the general question of when to make the effort of attempting to identify these complex types of interactions. Even with modest effect sizes, if a G×E finding is sufficiently replicated in human populations and supported by other experimental data, this information could provide insights into possible disease mechanisms. Finally, given the reduced power to detect G×E combinations with present methods, approaches that examine higher-order interactions should be taken on cautiously.

Despite many recent advances in analytical methods for G×E discovery, and some validation in recent years, additional statistical methods are needed for studies of copy number and rare genetic variation, survival traits, analysis of trios, and meta-analysis and pooling in large consortiums. In addition, many of the assumptions about expected G×E findings are based on results from genetic simulation studies, but these expectations have not always directly correlated to G×E observations in real population studies. Therefore, the question remains whether simulation studies have been designed with realistic assumptions about the underlying genetic architecture of the traits and whether better simulation approaches are needed (124).

Extended collaboration and data sharing will also advance G×E research. Large epidemiologic consortiums, such as the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) Consortium, with longitudinal measures of environmental exposures have been heavily leveraged in recent years as a way to examine repeated environmental exposures over time and attempt to incorporate cumulative and time-varying exposures into assessments of complex disease risk (125). There is also a need for further collaboration to allow validation of biomarkers in larger cohorts. Meta-analysis and pooling methodology and efforts will likely need to be advanced to have the power to detect G×E in rarer diseases. Standards are needed to describe the adequate criteria for identifying, reproducing, and reporting a G×E finding; a place to publish negative findings would allow researchers to avoid repeating failed experiments (11). There is also a need for greater integration and education with other fields to better design studies of G×E. Specifically, toxicology expertise will be needed to allow validation in experimental models of G×E discoveries. Last, the sharing of environmental and epidemiologic data has lagged behind genomic data sharing. Some have suggested that an environmental data-sharing policy mirroring the National Institutes of Health genomic data-sharing policy could advance this effort in the environmental health-science fields. However, there are unique sensitivities and ethical issues related to the sharing of environmental data that must be considered, including participant confidentiality and privacy issues (because environmental exposure data with global-positioning-systems information can allow specific identification of the sources of exposure) and legal/regulatory matters (e.g., regulatory reporting, remediation, and reform).

Researchers are exploring ways to apply G×E findings to risk-prediction studies as a possibility for targeted screening or intervention. Questions remain about the optimal approaches for risk-prediction models, including how to integrate biomarkers and external exposures and how best to model the joint effects of genetic markers, biomarkers, and lifestyle and environmental exposures (126). Although most statistical methods for detecting G×E focus on identifying departures from a multiplicative relative-risk model, the absence of multiplicative interactions will typically imply the presence of additive interaction (i.e., when there are marginal genetic and environmental effects). Additive interactions may have public health implications because they suggest that the difference in absolute risks between exposed and unexposed groups differs across genetically defined subgroups (105–108, 126). If an exposure causes disease, then an intervention to remove the exposure will prevent more cases in a genetically sensitive population than an in an equivalently sized genetically insensitive population. Important challenges that remain include determining whether the exposure in fact causes disease, developing effective interventions to change exposures, and evaluating whether targeted or population-level interventions optimize the risk-benefit trade-off. As with main effects, where it is well understood that observational findings of associations across individuals do not necessarily imply that an intervention to change exposure will change any individual's outcome, so an additive interaction does not necessarily imply that a genetically targeted intervention would be a more effective strategy for prevention. Modern methods of causal inference (127, 128) may be useful for estimating the causal difference in disease rates between genetically targeted and population-wide exposure interventions. Finally, the lessons and approaches for research into how the combination of genes and environment contribute to disease relates broadly to the studies of precision medicine and precision prevention. These types of studies may lead to insights for targeting prevention, intervention, or treatment in the future.