Gene-environment interactions in human health

Abstract

Gene-environment interactions (G × E), the interplay of genetic variation with environmental factors, have a pivotal impact on human complex traits and diseases. Statistically, G × E can be assessed by determining the deviation from expectation of predictive models based solely on the phenotypic effects of genetics or environmental exposures. Despite the unprecedented, widespread and diverse use of G × E analytical frameworks, heterogeneity in their application and reporting hinders their applicability in public health. In this Review, we discuss study design considerations as well as G × E analytical frameworks to assess polygenic liability dependent on the environment, to identify specific genetic variants exhibiting G × E, and to characterize environmental context for these dynamics. We conclude with recommendations to address the most common challenges and pitfalls in the conceptualization, methodology and reporting of G × E studies, as well as future directions.

Introduction

Genome-wide association studies (GWAS) have contributed to the identification of genetic loci associated with complex traits or diseases far beyond the scope provided by candidate gene association or family linkage studies. However, for most complex traits, broad-sense heritability $\left(H^2\right)$ estimates are significantly higher than the heritability attributable to the additive combination of effects from SNPs, that is, SNP-based narrow-sense heritability $\left({h^2}_{\text{SNP}}\right)$ — a phenomenon known as “missing heritability”¹. For example, asthma ${h^2}_{\text{SNP}}(9-33\%)$ estimates are consistently lower than $H^2$ estimates from family-based studies (55–90%), with larger estimates for childhood-onset than adult-onset asthma². In addition, despite the high genetic correlation between asthma subtypes (r_g = 0.78), genetic effects are greater in childhood-onset than in adult-onset asthma³. Together with differences in disease manifestation and triggers of asthma exacerbations⁴ across ages of onset, these discrepancies suggest the presence of interactions between genes and age, which are potentially driven by environmental factors, with heterogeneous genetic effects on asthma across the life course.

Recent computational advances and the increasing availability of large-scale genomic datasets for diverse populations, combined with deep phenotype information from accessible electronic medical records and geo-coded environmental data, offer unparalleled opportunities to explore how genetic predispositions interact with environmental exposures, ultimately advancing our understanding of complex traits and diseases. Although other factors may contribute to the missing heritability — such as unaccounted, rare or structural variation, non-linear genetic effects, epigenetic inheritance, horizontal and vertical microbial transmissions or underestimation of the shared environment in twin studies^1,5 — gene-environment interactions (G × E) thus merit further exploration.

G × E are statistically defined as a departure from predictive models that consider the phenotypic effect of genetics and environmental exposures alone (Fig. 1). Investigating G × E can lead to the identification of novel genes that do not exhibit marginal effects⁶ to provide insights into underlying biological mechanisms, strengthen causal inference derived from observational studies⁷, affect public health policy regarding prevalent environmental exposures and refine prevention programmes by building better prognostic models^8–10. However, genome-wide screenings of G × E, also known as genome-wide interaction studies (GWIS), face multiple challenges that require consideration^8,9,11. For example, these analyses require larger sample sizes to achieve adequate statistical power to detect G × E effects compared with main genetic effects. Additionally, it can be difficult to appropriately model the complexity of G × E dynamics, especially with the statistical and computational burden of multiple testing.

Fig. 1 |. G × E modelling.

A, The interplay between genetic and exposure risk factors on a phenotype can be depicted through traditional epidemiological models of gene-environment interactions (G × E), with variable genetic effects across environments. Model scenarios include situations where the genotype exacerbates the exposure effect on the phenotype, but there is no genetic effect in non-exposed individuals (Aa); the exposure factor exacerbates the genetic effect on the phenotype, but has no effect on the phenotype in the absence of the risk genotype (Ab); both the genotype and exposure factors are required to increase the risk (Ac); and either the genotype or the exposure factor can influence risk, with their combined effect potentially differing from their individual effects (Ad). B, G × E on a phenotype. The model includes the effect of two genotypes (G0, G1) and two environments (E0, E1) on the phenotype. However, genetic and exposure factors may not necessarily interplay (Ba). When interactions occur, non-crossover interaction effects may present in either or both the additive (Bb) or multiplicative scale (Bc). Given that additive interaction suggests a differential absolute risk reduction associated with one factor across different levels of another, it can identify subgroups that would benefit the most from public health interventions. Conversely, multiplicative interactions imply that the combined effect of risk factors diverges concerning the product of individual effect and provide a better fit for the joint G × E effects⁷⁸. Although less common, crossover interaction effects (Bd) can alter the genotype rank order based on environmental influences, as seen in flip-flop effects for ABCA1 and ATP8A1 loci and environmental tobacco smoke exposure on bronchial hyper-responsiveness¹²¹. C, Mechanisms underlying interaction with environmental exposures when genetic variation, DNA methylation and gene expression levels are considered. G × E models can be expanded to capture regulatory effects of genetic variation on different biological layers, such as DNA methylation and gene expression levels¹³⁰. Simultaneously, these models can encompass environmental effects across biological layers. This multidimensionality allows for a more comprehensive understanding of how genetic and environmental factors interact and influence various aspects of biological processes. Part A adapted with permission from ref. 148, Elsevier. Part B adapted from ref. 149, Springer Nature Limited.

The molecular mechanisms underlying environmental influences on human health have been recently reviewed elsewhere^11–13. Here, we provide guidance for designing G × E studies depending on the research question and available data. Furthermore, we discuss G × E methods for analysing single and polygenic G × E, with a specific focus on elucidating the genetic contributions to G × E and G × E detection and characterization. We also provide an overview of the current landscape of G × E findings for human complex traits and diseases, and highlight considerations for future research.

Designing a G × E study

Environmental, socio-cultural and genetic factors dynamically affect human health across the lifespan by influencing biological responses at the molecular, cellular or systemic levels, shaping behavioural responses and processes. This complex interplay unfolds across multiple levels of socio-ecological dimensions, spanning from the individual level to a broader, global environment (Fig. 2). Thus, the design of a G × E study is an iterative process between choosing an appropriate research question and analytical approach, and the selection, cleaning and harmonization of available genetic data and environmental data.

Fig. 2 |. A framework depicting the joint contribution of genetic susceptibility and environmental and social exposures to human health across the lifespan.

As an extension of the Integrated Socio-Environmental Model of Health and Well-Being (ISEM)¹⁵⁰, this model illustrates how intra-individual, community-level, social and physical environmental factors contribute distinctly and collectively to shaping human health. Individuals are exposed to and affected by a diverse range of interconnected physical environmental factors, including natural elements, built-in environments and global conditions, as well as social-structural factors. This dynamic interplay is further influenced by age, culture and policy. This intricate relationship triggers behavioural and biological responses that affect health across various levels, encompassing behavioural outcomes, system and identity, as well as biological responses at systemic, multi-organ, cellular, subcellular and molecular levels. These responses may occur independently or interact, underscoring the complex interplay between biology, behaviour, environment and society. Adapted from ref. 150, Springer Nature Limited.

Choosing a G × E analytical framework

Understanding how genetics and environmental contexts interact to influence human health hinges on central strategies (Fig. 3) that require specific methodological approaches to address different underlying questions (Table 1). A key question is whether polygenic G × E contribute to a human trait (and by how much) across one or more environmental contexts. For this purpose, genetic variants are considered in aggregate through population-level parameters, which characterize genetic liability and its relationships with environment across a study population (that is, G × E heritability $\left({h^2}_{\text{GE}}\right)$ and G × E correlation $\left(r_{\text{GE}}\right)$), or individual-level factors, which capture genetic risk per individual (that is, the polygenic score (PGS)). Differences in genetic effects per environment can also be leveraged to strengthen causal inference using Mendelian randomization (MR).

Fig. 3 |. General overview of G × E methods.

Data for gene-environment interactions (G × E) studies can be linked through multi-stream sources. To ensure suitability for analysis, rigorous quality control, harmonization and normalization procedures are essential, particularly when data are collected from different study sites. Screening approaches involve one or multiple of the approaches shown. The genome-wide association study (GWAS) and genome-wide interaction study (GWIS) analysis equations depicted assume binary outcomes $(Y)$ for simplicity. The most common approach models G × E using a generalized linear model, where the probability $(Pr)$ is calculated given that outcome ${\beta }_0$ is the intercept, ${\beta }_{\text{G}}$ is the genetic (G) main effect, ${\beta }_{\text{E}}$ is the exposure (E) main effect, ${\beta }_{\text{G}\times \text{E}}$ is the G × E effect and ${\beta }_{\text{C}}$ is the covariate(s) (C) effect (Box 1). The model tests whether there is a significant interaction effect between genetics and environment, in which the null hypothesis assumes no interaction between genetics and environment $\left({\text{H}}_0:{\beta }_{\text{G}\times \text{E}}=0\right)$ and the alternative hypothesis suggests a non-zero interaction $\left({\text{H}}_1:{\text{β}}_{\text{G}\times \text{E}}\ne 0\right)$. In a case-only design, the equation estimates the log odds of specific genetic factors $(G=\mathrm{g})$ given exposure and covariates, while accounting for the individual being a case $(Y=1)$. The exposure-stratified design divides the study population into subgroups based on the environmental exposure factor (where $E=0$ represents unexposed and $E=1$ represents exposed). Within each subgroup, a regression model is fitted to estimate the probability of the outcome given genetic factors and covariates. $\text{γ}$, effect of the interaction between variables on the outcome; 1DF, one degree of freedom; 2DF, two degrees of freedom; ${h^2}_{\text{GE}}$, narrow-sense heritability attributable to G × E; $I_{\text{GZ}}$, gene-environment interaction; IPGS, interaction polygenic score; M, total number of variants; MR, Mendelian randomization; PGS, polygenic score; PGS × E, polygenic score-environment interaction; $r_{\text{GE}}$, G × E correlation; $U$, unmeasured confounder; vQTL, variance quantitative trait locus; vPGS, variance polygenic score; $Z$, interaction covariate.

Table 1 |.

G × E study designs and methods

Application	Approach	Advantages	Limitations
Traditional common study designs a
Screening	Case-control	Modest sample size requirements for rare diseases; can individually match on confounders	Recall bias for environment; difficulties selecting proper controls and higher cost than case-only
Screening and characterization	Case-only approach	More cost-efficient, simpler to implement when recruiting suitable controls is challenging and more precise effect size estimation compared with case-control	Requires gene-environment independence; provides estimates for G × E, and not main genetic and environment effects; biased estimates under population stratification
Family-based designs a
Screening	Case-parents	Assessment of main genetic, G × E and joint effects; protects against confounding due to population stratification (by matching genetic background with case-parents); more powerful than case-control studies for G × E	Often infeasible for late-onset diseases; possible bias in G × E if genetics and environment are associated within parental mating types; difficult to enrol complete triads; unsuitable for environment effect estimates; may require additional control for exposure-related genetic stratification
	Case-siblings	Protects against confounding due to population stratification (matching genetic background between cases and controls); can be applied in late-onset disease	Enrolment of discordant sibships can be challenging; overmatching for genetic main effects
GWIS-like designs a
Screening	Several designs	Several analytical frameworks available; offer marked versatility for subsequent genetic and functional analyses, and facilitate replication and portability assessment; primary goal is screening, yet summary statistics and genetic data can also be used to characterize Gx E liability and for prediction	Biased to detect common variants with modest to small effect size in non-coding regions; requires large sample size for adequate statistical power to detect associations
Analytical frameworks
Screening	2DF test	Boosts power to detect signals by testing main genetic and G × E together	Requires 1DF to disentangle whether findings are driven by genetic main effects, G × E effects or both; inflated type I error rates in presence of unaccounted covariate-environment confounding, when compared with meta-regression; potential bias on G × E effects under heterogeneity of main genetic effects and proportion of exposed individuals across studies; potential noisy main genetic effect estimates for continuous exposures
	Screening approaches	Computational and statistical burden of multiple testing from genome-wide scans; can be combined with aggregate-level effects, PGS and/or quantitative trait variability	Limited applicability for G × E screening using marginal effects loci; can miss some interactions
	Set-based methods	Aggregation of weak individual effects can identify signals undetectable in isolation (indispensable for rare variants); different set-based or hybrid methods boost statistical power to detect signals; multi-exposure and multi-phenotype frameworks available	Statistical power fluctuates based on underlying method in the presence of heterogeneous effects and varying numbers of causal variants
Screening and characterization	1DF test	Discern whether findings are driven by main genetic, G × E and joint effects	Lower power to identify variants with moderate joint effects and low genetic main effects compared with 2DF test
	Exposure-stratified framework	Characterizes main genetic effects across environmental strata (concordant, single and opposite effect direction)
	Metaregression	Applicable to quantitative traits; tests main genetic, G × E and joint effects; robust to covariate-environment confounder; accommodates overlapping data and heterogeneity	No widespread application to date
	Machine learning	Can outperform regression-based analysis and accommodates non-linear effects	High computational burden for genome-wide analysis; high dimensionality and noise reduction required; limited toolset
	Bayesian approaches	Accounts for uncertainty of gene-environment independence assumption and can outperform the case-only approach	High computational burden for genome-wide analysis (in some cases); limited toolset
	GEVACO	Tests the joint genetic and dynamic G × E with continuous traits; applicable to a single marker and sets of variants	Currently unavailable for categorical phenotypes
	Multilevel modelling	Applicable to longitudinal data; accounts for kinship; yields higher statistical power than linear mixed models and regression analyses; robust to partially missing data	Requires strong assumptions for causal inference; requires sufficient representation across level units with dataset
	Heritability	Focus on polygenic G × E; can be used for G × E screening under varying genetic variance	Potentially limited contribution to complex traits; unclear how environment-partitioned and cell or tissue-partitioned heritability contribute to human health
	Correlation	Focus on polygenic G × E; can be used for G × E screening under imperfect genetic correlation and prioritize trait-environment pairs with substantial trait variance for G × E screening	Presence of r GE may act as a confounder when modelling G × E
G × E liability and characterization	MR	Applicable to individual-level or summary statistics; novel approaches leveraging environmental data expected to emerge	Limited toolset; software-specific limitations in application for individual-level and summary statistics analyses
G × E liability, screening and prediction	PGS×E	Focus on polygenic G × E; can be used for G × E screening under amplification and varying genetic variance	Limited accommodation of antagonistic effects; noisy PGS×E estimates when main effects PGS considered; requires disentangling main genetic and G × E effects, potential confounders and rGE; unclear how environment-partitioned and cell or tissue-partitioned PGS contributes to human health
Screening and prediction	Quantitative variance genetic variability	Adaptable for polygenic G × E using vPGS from vQTL; capture outcome plasticity better than traditional PGS; applicable to large datasets with incomplete or lack of environment data; novel approaches leveraging environment data expected	Potentially limited number of vQTL and/or small contribution to complex traits; immature statistical methods; lack of frameworks to discern phenotypic dispersion, epistasis, effects on phenotype distribution and phantom vQTL; lack of frameworks for within-individual variability or diverse populations

1DF, one degree of freedom; 2DF, two degrees of freedom; G × E, gene-environment interactions; GWIS, genome-wide interaction study; MR, Mendelian randomization; PGS, polygenic score; PGS×E, polygenic score-environment interaction; r_ge, G × E correlation; vQTL, variance quantitative trait locus; vPGS, variance polygenic score.

^aFor a comprehensive summary of study designs, we refer readers to Table 1 in ref. 10.

The avenue of investigation entails identifying the specific genetic variants implicated in G × E and characterizing the dynamics of these genetic variants with environmental factors. Analytical approaches for single-marker and polygenic analysis, including the PGS, can offer insights into the individual-level variability in genetic risk with environment. Novel G × E screening methods have investigated the variance of continuous phenotypes across loci^14–16 using variance quantitative trait locus (vQTL) or variance polygenic score (vPGS) analyses in biobanks (Fig. 3).

Data collection and quality control for G × E studies

The advent of microarrays and affordable sequencing, coupled with increased computational capacity to analyse genetic data, has paved the way for GWIS to supersede candidate gene strategies. Testing for interaction of genetic variation across the genome and phenotype(s) of interest, GWIS offer greater versatility for downstream genetic and functional analyses, as well as improved capabilities for replication and portability assessment. Thus, it is now standard practice to conduct multi-study, large-scale genomic analyses to boost statistical power to detect genetic signals. This approach requires appropriate quality control practices to control for potential batch or platform effects through common filters and, if needed, genome-wide imputation using appropriate reference panels¹⁷. Although GWIS involve similar steps to GWAS, including data collection, genotyping or sequencing, quality control, imputation (if needed), association testing, meta-analysis and replication, additional considerations are necessary to ensure the integrity and harmonization of genetic, phenotype and environmental data (Fig. 3).

Issues arising from inadequate control of heterogeneity by data collection site are amplified in environmental and outcome data, as well as during biospecimen collection for their measurement, and can affect downstream analyses when combining and comparing data from multi-centre studies or multiple individual studies^18,19. The PhenXToolkit²⁰ is an important tool for data harmonization, as it provides protocols and definitions for a range of variables over 30 domains, including demographics, social determinants of health and numerous biological systems, such as respiratory or cardiovascular systems. Missing phenotype or environmental exposure data can be imputed, leveraging the correlation structure of the available phenotypes, environmental data or family structure either from the same study or external sources^21–24. When imputation is unviable (for example, insufficient data across strata or multiple missing variables), it has been shown that considering only environmental data that are comprehensively available for unexposed individuals while allowing for missingness in exposed individuals outperforms any other scenario²⁵.

In planning a multi-study analysis, the comparison of questionnaire data and biological assessments across pre-existing studies for harmonization facilitates the linkage of study variables, enabling prioritization of specific variables for meta-analysis¹⁸. In addition, power calculations help comprehend the constraints of the research design given available data, identifying prerequisites for new studies and determining appropriate analytical methods. Thus, they should be routinely considered in the conceptualization stage and can be easily performed using available software^26–31.

Researchers should be aware of the impact of trait normalization and scaling on the interpretation of their findings. The choice of measurement scale, such as raw or log scales, influences the magnitude of interaction effects. Although trait transformation can reduce scaling effects and accommodate non-normal distributions, it can also generate biases on the null hypothesis testing³². Moreover, scale dependency can result in inaccurate quantification of interaction effects, potentially leading to false positives or negatives³². Therefore, when environmental data across studies vary in scale, researchers could consider outlier studies for validation¹⁸, across geographic locations, over time or life stages, with careful reporting and interpretation of the findings.

Overall, careful consideration of data selection, cleaning and harmonization must be conducted for G × E analyses, as potential issues with heterogeneity are amplified with non-genetic data, depending on the specific research question and analytical decisions. Moreover, when pooling studies, downstream quality control at the study or post-meta-analysis level is necessary to avoid analytical bias. A common issue for G × E studies arises from insufficient control for covariation between the covariates and the environment, which could be mitigated by assessing model misspecifications before genome-wide scans³³ or applying exposure-stratified analysis³⁴. Furthermore, available frameworks can facilitate summary statistics harmonization and assessment of genomic control, potential analytical issues and cross-study heterogeneity^17,35. In addition to applying filters based on the minor allele count and imputation quality¹⁷, the study-specific choice of approximate degrees of freedom — estimated as the product of these metrics — is essential for eliminating low-informative variants³⁴.

Statistical G × E methods

Statistical methods are applied to systematically test for interactions between genetic variants and environmental factors. This involves testing each variant (or set of variants)-environment interaction using different statistical approaches while controlling for potential confounders and correcting for multiple testing (Table 1 and Fig. 3). Apart from considering the most adequate analytical framework(s) to approach the research question, thorough consideration of the anticipated genetic model of the trait and identification of potential sources of analytical bias are necessary to ensure a nuanced interpretation of findings. Although statistical interactions can reflect mechanistic interactions, they do not inherently imply biological or functional interaction where two exposures physically interact. The choice of statistical method should align with the study rationale, focusing on either or both the additive or multiplicative scale (Fig. 1), or exploring exposure rate or case status variation within specific genetic subgroups. Therefore, reporting choices and assumptions for a study’s statistical modelling, as well as genetic effects for all tested scenarios, is crucial for contextualization and replication assessment.

Determining differences in polygenic liability by exposure

A pivotal question in G × E research revolves around quantifying the influence of the environment on genetic liability. These following methodologies prove valuable in causal inference, distinguishing parental genetic and environmental effects, providing insights into trait aetiology and/or prioritizing environmental exposure-outcome pairs for further refinement^36,37.

Quantifying differing genetic contributions by environment at the population level.

The extent to which G × E contribute to the proportion of phenotypic variation of human traits $\left({h^2}_{\text{GE}}\right)$ remains unclear and is likely highly dependent on population, context and specificity of trait. Although ${h^2}_{\text{GE}}$ is generally considered to be low^38–42, it exhibits heterogeneity in estimates based on the analytical approaches^38–41,43. However, exposure-partitioned and functionally partitioned heritability analyses suggest variable genetic influence by exposure strata (Box 1). Most analytical tools currently estimate ${h^2}_{\text{GE}}$ for a single exposure, but there is growing interest in the overall intersectional contribution of multiple environmental contexts to human health^40,42,44 (discussed below). Heritability is often calculated using GREML-based^40,41,45,46 methods, which require individual genotypes and yield reliable estimates. Conversely, linkage disequilibrium score regression^4,48 gains computational efficiency by relying on summary statistics, and therefore may be more accessible.

Box 1. Novel advances on G × E by the largest multi-ancestry GWIS meta-analysis.

The CHARGE Gene-Lifestyle Interactions Working Group recently leveraged 28 genome-wide interaction study (GWIS) summary statistics from diverse individuals¹³¹ to provide methodological advances and characterize gene-lifestyle interactions on lipid levels and blood pressure traits. Their comprehensive gene-environment interactions (G × E) framework showed that the two degrees of freedom (2DF) joint test systematically boosts the statistical power to detect significant loci for ancestry-specific and trans-ancestry analyses compared with the one degree of freedom (1DF) test. However, future methods should consider the low correlation between marginal and interaction genetic effects observed within the identified 2DF loci. Along the same line, they also evidenced the limited applicability of two-step approaches reliant on marginal genetic effects for G × E screening.

Their analysis of the phenotypic variance explained by marginal genetic effects and interaction effects implied that the contribution of interaction effects in addition to marginal genetic effects is relatively limited. Despite this, their work revealed heterogeneity in heritability partitioned by exposure. Further stratification of exposure-stratified heritability by functional and cell-type annotations revealed differential variability between exposed and unexposed strata, suggesting specific biological mechanisms at play. Although these differences in exposure-stratified heritability may arise from unknown biases, further analysis are required to disentangle the contribution of stratified analyses to G × E liability, characterization and prediction.

The CHARGE Gene-Lifestyle Interactions Working Group research also suggested potential ancestry-specific patterns of gene-lifestyle interactions, with differences in phenotypic variability explained by interaction effects, particularly among populations of African descent. This underscores the importance of conducting G × E research on diverse populations.

Genetic variation also shapes our experiences with the environment, such as individual-level lifestyle factors (that is, smoking) or community-level influences (that is, social support). $r_{\text{GE}}$ measures the extent to which genetics is linked with environmental context variation. Although $r_{\text{GE}}$ cannot be used for risk prediction at the individual level, it can be leveraged to distinguish G × E variants in a case-control setting⁴⁹ and prioritize trait-environment pairs with substantial trait variance for locus-specific analysis³⁷. The presence of $r_{\text{GE}}$ may act as a confounder when modelling G × E, leading to spurious results⁵⁰ and biased interaction estimates⁵¹. Appropriate study designs, such as family-based designs⁵², and analytical approaches encompassing joint modelling of interaction and correlation⁵³ can contribute to disentangle the specific contribution of each factor.

Lastly, another methodology to assess differences in genetic contributions to health outcomes by environmental context is MR. Rather than quantifying differential genetic liability by environment, MR uses genetic variants as instrumental variables to investigate the causal effects of a modifiable environmental factor on a health outcome. For instance, the influence of body mass index (BMI) on blood pressure could be investigated through BMI-associated variants, provided essential assumptions are met⁵⁴. One of the sources of MR estimate biases is the presence of horizontal pleiotropy, which occurs when a genetic variant is associated with a health outcome through biological pathways unrelated to the exposure of interest (for example, BMI-associated variants do affect blood pressure through additional confounder pathways). Strengthened MR tools^55–60 can leverage an environmental characteristic modulating the strength of the association of variants on the exposure to improve the precision of causal inference estimates. For example, MR-G × E⁵⁴ refined causal estimates for the effect of BMI on systolic blood pressure in the UK Biobank, by accounting for alcohol and physical activity as interaction covariates (represented as $I_{\text{GZ}}$ in Fig. 3).

Determining how environmental context alters distributions of genetic risk at the individual level.

PGSs are calculated for individuals by aggregating the weighted effects of multiple genetic variants from GWAS summary statistics to improve genetic risk prediction beyond individual genetic variants. Evaluating the interaction between the PGS and the environment enables the characterization of the polygenic Gx E relationship and allows for G × E screening. Significant PGS-trait pairs could prioritize traits for both single-marker and further multi-marker analyses. Additionally, they can aid in uncovering the underlying biology of the trait by assessing its association or interaction with different exposures, outcomes or omic data. They could also provide insights into the variation in genetic risk distribution by environmental context. For instance, the clinical relevance of exposure-stratified PGS estimation remains unclear at present given the differences in exposure-stratified heritability (Box 1). Nevertheless, GWAS-derived polygenic score-environment interaction (PGS × E) may be susceptible to bias due to small sample size measurement errors in the original GWAS and confounding by $r_{\text{GE}}$ if unaccounted for. In addition, PGS × E may fail to capture antagonistic G × E effects and overlook variants with G × E effects but minimal genetic main effects. Despite these limitations, the prevailing approaches in PGS × E development continue to rely on GWAS.

PGS × E estimation can be obtained using a case-only scenario under the assumption of gene-environment independence⁶¹ whereas the correlation between the PGS and environment can be discerned by joint modelling of interaction and correlation using a case-control design⁵³. GWIS summary statistics can also be used to derive variant weights based on the main effects for the genetic variant and interaction term for the genetic variant and the environmental exposure, or the interaction term alone, with contradictory findings regarding performance compared with the traditional PGS × E across studies^62–64. Given the sample size requirements for G × E, analytical strategies advocate for sample overlap between datasets^51,65. For instance, PIGEON⁵¹ showed that, under a variance component analytical framework, a genetic correlation analysis of GWIS and GWAS summary statistics provided unbiased G × E estimates compared with PGS × E.

Identification of G × E relationships at variant and set levels

With the rise of next-generation sequencing data, traditional epidemiological study designs¹⁰, such as case-only or family-based studies (Table 1), are shifting to GWIS-like approaches in unrelated individuals from diverse populations through consortia or biobanks. Instead of characterizing overall differences in genetic influences on human health by environmental context, these analyses aim to estimate how the environment can influence the association between specific genetic variant(s) and outcome. Here, we outline common strategies for single-marker and multi-marker (or set-based) testing for G × E screening and characterization while ensuring computational tractability.

Variant-level effects assessed through joint or stratified approaches.

A model approach to estimate G × E can take the form of a joint model, in which a single statistical model is constructed that includes both main effects for the genetic variant and the environmental exposure as well as an interaction term for the genetic variant and the environmental exposure. Alternatively, in the stratified approach, the study population is divided into subgroups based on different environmental exposure levels, and genetic effects are estimated separately within each stratum.

The most common approach has G × E modelled using a generalized linear model (Fig. 3), including terms for the main genetic and environmental effects and an interaction term for the genetic and environmental effect. The model tests whether there is a significant interaction effect between genetics and environment in which the null hypothesis assumes no interaction between genetics and environment, and the alternative hypothesis suggests a non-zero interaction. The one degree of freedom (1DF) approach examines differences in genetic effects on an outcome given an exposure, whereas the two degrees of freedom (2DF) approach jointly tests⁶⁶ whether the combined impact of genetic variation and exposure on outcomes differs from their individual effects (Box 1). In the joint model, a 1DF and 2DF inverse-variance weighted meta-analysis can be used for the interaction term⁶⁷ and joint main genetic and interaction effects⁶⁸, respectively. Although the 2DF test may have better power for signal detection⁶⁶, conducting 1DF testing is essential to clarify whether the signal is driven by main genetic effects, interaction effects or the combination of both (Table 1).

An alternative approach tests main genetic effects stratified by outcome or exposure strata. Under the assumption of gene-environment independence, the case-only analysis provides unbiased effect size estimates for the interaction term in affected individuals (Fig. 3). Alternatively, the exposure-stratified framework can distinguish opposite effects from concordant or single-stratum effects through stratumdifference⁶⁹ or stratum-specific scans. To combine the stratified effects, a 1DF meta-analysis can account for between-strata relatedness⁷⁰. Under the absence of stratified data, the J2S framework⁷¹ estimates the stratified and marginal effects from joint G × E summary statistics. Overall, the joint 2DF meta-analysis outperforms the stratified framework, particularly for low-frequency variants⁷². However, it may lead to a higher type I error rate under covariate-environment confounding³⁴. To address this issue in quantitative traits, meta-regression evaluates the main genetic effects from groups stratified according to exposure strata³⁴. By pooling multiple studies, meta-regression has uncovered sex-specific and age-specific effects for anthropometric traits in Europeans⁷³. Moreover, several meta-analysis frameworks have been extended to accommodate random effects, overlapping data or multiple phenotypes^74–76.

Approaches to increase statistical power for G × E.

Analyses investigating the role of both genes and environment on human health often have reduced statistical power compared with standard GWAS approaches. Two contrasting approaches can increase statistical power by either screening loci for follow-up to conduct fewer comparisons or combining variants and analysing them jointly.

Screening approaches prioritizing sets of variants or genes for G × E analysis using data-driven or hypothesis-driven approaches can diminish the computational and statistical burden of multiple testing from genome-wide scans. To increase the statistical power to detect interactions, these strategies leverage aggregate-level effects, the PGS and/or quantitative trait variability. Some methods lacking computational tractability for genome-wide screening are highly valuable for modelling G × E for prioritized variant sets, such as machine learning⁷⁷ and Bayes approaches^78,79 (Table 1). The data-driven strategies often implement two-step approaches, with genome-wide scans for genetic main effects followed by interaction testing^78,80–83 (Box 1). By contrast, hypothesis-driven strategies rely on biological and functional annotations to prioritize gene or variant sets for interaction testing. For instance, candidate genes could be selected based on their pharmaceutical targetability or implication in xenobiotic metabolism using the ComparativeToxicogenomicsDatabase⁸⁴ and the Drug-GeneInteractionDatabase⁸⁵.

A novel screening method for large-scale studies incorporates screening for individual loci or PGS influencing quantitative trait variability, followed by interaction testing for environmental factors. vQTL-environment interaction (vQTL × E) screening has been applied to study interactions on human traits in the UK Biobank, including anthropometric and cardiometabolic traits^14–16. For example, the vQTL rs1229984 (ADH1B) evidenced a positive association between alcohol intake and alanine transaminase levels in CC carriers but was absent in TT carriers¹⁵. Despite the enhanced ability of vQTL effects to capture the plasticity of G × E dynamics compared with traditional GWAS variant weights⁸⁶, their contribution to certain traits is limited¹⁶ and analytical challenges remain (Table 1). Spurious vQTLs may arise from phantom effects stemming from untagged causal additive QTL variants, nonlinear trait-environment relationships or scale artefacts⁸⁷. Mitigating these effects may involve sensitivity analyses and/or replicating effects on both additive and multiplicative scales. In the context of pervasive heterogeneity of pleiotropic effects on the additive scale for the outcome — where outcome heteroscedasticity is probable — some MR methods^54,56,59 become bias-prone, prompting consideration of alternative frameworks⁵⁷.

In contrast to decreasing the scope of variants investigated, set-based analyses amplify the statistical power for detecting G × E signals by aggregating individual variant effects within genes, regions, pathways or the genome. The aggregated effect of multiple weak signals, possibly undetectable individually, is particularly relevant for rare variant identification. In the presence of synergistic G × E effects, set-based tests outperform other methods⁸⁸. Several types of set-based G × E have been developed based on the variance component^89,90, trait similarity regression^91,92, the burden test and sequence kernel association test^93–97, optimal weighting⁹⁸, adaptive combination of Bayes factors⁸⁰, machine learning⁹⁹ or a combination of these^100–104. Some of these methods accommodate multi-trait analysis of rare variation¹⁰⁵, relatedness¹⁰⁰ or joint tests^100,102, and are available for longitudinal¹⁰⁶ or large-scale biobank data^89,100,107. Using whole-exome sequencing data from the UK Biobank, a set-based analysis revealed a significant MC4R by sex interaction on BMI, consistent with prior research on sex-specific MC4R effects on brain structure and eating behaviour¹⁰⁰. Nevertheless, the landscape of statistical frameworks for meta-analysis of set-based G × E and joint test summary statistics is limited to fixed-effects and random-effects meta-analysis¹⁰⁸, additional filtering prior meta-analysis¹⁰⁹, variance component methods¹¹⁰ and linear mixed model-based meta-analysis¹¹¹.

Considerations for large-scale biobanks.

The inception of large population biobanks will foreseeably lead to the development of numerous computationally tractable G × E statistical frameworks for large-scale genomic analysis. A few are detailed here that account for various potential setbacks when considering large-scale biobanks and their data with respect to ascertainment and availability. Given the outcome-independent recruitment designs, biobanks often have unbalanced case-control ratios, for which saddle-point approximation has been tailored for GWIS¹¹². Moreover, fastGWA-GE conducts genome-wide testing for main genetic, G × E and joint effects while accounting for relatedness either through pedigree information or a sparse genetic relationship matrix¹¹³.

Aggregate-level effects can be assessed in large sequencing studies using different methods^{89,100,101,104,107}. The state-of-the-art approximate method, MAGEE¹⁰⁰, implements several interaction and joint tests while accommodating relatedness using a genetic relationship matrix. Other methods have been proposed to boost statistical power by modelling multiple factors that may exert antagonistic or synergistic effects, such as StructLMM⁴⁴ or LEMMA⁴⁰. StructLMM⁴⁴ models environmental similarity between individuals as a random effect and adopts a variance component test score for G × E testing⁴⁰. By contrast, LEMMA is a Bayesian whole-genome approach that jointly models the main genetic and G × E effects considering an environmental score aggregating the effect of multiple environmental exposures⁴⁰.

Plasticity of G × E dynamics

Unravelling the complexity of G × E (Fig. 2) requires attention to nuanced relationships, with challenges arising from dynamic interactions among multiple genetic and environmental factors. This influence of time becomes vital when understanding G × E, both within the study design process as well as interpretation of results. Genetic effects can exhibit context specificity, whereby the environment can modulate the effect’s magnitude or even invert it based on the individual’s personal environmental context across the lifespan and interactions across the exposures. This plasticity requires multi-exposure, multi-trait or environmental score approaches to account for complex dynamics; such approaches have been popularized as a means of leveraging pleiotropy to identify G × E signals^{40,44,83,114–116}. These strategies offer enhanced robustness and power compared with standard single environmental exposure analysis, potentially revealing novel loci^{44,88,116,117}.

Moreover, the endeavour to model the plasticity of the G × E relationship also stems from the variability of the interaction across contexts, such as differential magnified effects across environmental strata that can be captured by the meta-regression approach³⁴. Alternatively, semi-parametric varying-coefficient joint testing enables the identification of signals missed by the traditional linear effects tests¹¹⁸. As an example, multilevel models using time-varying terms explored the impact of resilience and trauma on genetic effects in war-related psychosocial stress over 48 weeks among Syrian refugee and Jordanian non-refugee youth¹¹⁹. The study revealed significant interactions between resilience and 5-HTTLPR and COMT variants, but no time-related changes were observed.

Family-based studies of inherited effects

Family-based designs for G × E, more robust against population structure than case-control methods, are limited^120,121, possibly due to challenges in replication assessment arising from disparities in ascertainment, study variable definitions and environmental exposure distribution. Nevertheless, these designs prove convenient for exploring the causal environmental influences of parental traits on offspring outcomes, as the PGS can capture both genetic and environmental factors. Genetic transmission and genetic nurture on child outcomes have been investigated using parental and offspring PGSs^122,123. Another method enriched their structural equation models with information about transmission of parental PGSs to explore the effect of parental genetic and environmental influences on offspring trait variation while considering assortative mating⁵⁸.

Challenges and future directions

The investigation of G × E faces methodological and conceptual challenges related to statistical power, exposure/omics data measurement and quality control, study-specific confounders, co-linearity of environmental exposures, genetic ancestry and G × E dynamics. Table 2 summarizes our recommendations for the conceptualization, methodology, reporting and sharing of future G × E research, expanding on those from Dunn et al.¹²⁴.

Table 2 |.

Recommendations for the conceptualization, methodology, reporting and sharing of G × E research

Item	Recommendation
Conceptualization
Aim	Identify strategic opportunities for systematic assessment of environmental and outcome data in pre-existing studies; define the relevant area(s) of research — G × E liability, screening or characterization; define the relevant catalogue of genetic variation (for example, SNPs, CNV and so on) and environmental exposure(s) to assess
Plasticity of G × E	Consider the most appropriate strategies to address the G × E dynamic in your research question based on frequency and duration of exposure; testing of linear, non-linear or non-constant effects; and multi-trait, multi-exposure or ‘exposome’ effects
Sample size and power	Conduct power calculations to quantify potential bias and implement correction measures wherever possible (useful references26–31)
Multi-omics	Define whether integration with other omic data may provide additional relevant insights into your research question
Subgroup analysis	Consider subgroup analysis in the main stage or as a sensitivity analysis, and the suitable subgroup stratification strategies to address the research question
Replication	Aim to assess your findings for replication and to validate previous associations where possible; when the same environmental measurements are not available across studies, consider proxies for replication across geographic locations, over time or life stage
Methods
Study design	Report on the study design (including type of study, power calculations, sample size), advantages and limitations to address the research question
Setting	Describe the locations and periods of recruitment, exposure and follow-up
Sample	Include descriptive statistics of the whole sample and sample subgroups (for example, exposed versus non-exposed) for all the considered variables
	Report on the assessment of genetic ancestry, the ethnicity/ancestry group categorization, and the advantages and limitations of the selected approach (useful references142,143)
Variables	Define all outcomes, exposures and potential confounders; define cross-study harmonization strategy, if applicable
Data sources or measurement	Report sources of data and methodological details for assessing each variable of interest in the report; report the process of data harmonization, as well as choice and assumptions for data transformation and scaling; describe the comparability of assessment methods if multiple methods or groups were considered; describe how repeated measurements were processed, if considered; describe how missing data were handled
Genetic data	Follow standard profiling and quality control practices and incorporate additional procedures as needed (useful references140,144,145)
Statistical analysis
Reporting of model selection	Report on the reasoning and assumptions for model selection
Confounding	Include all the relevant covariates in the analysis and report how covariates were deemed relevant; evaluate model misspecifications; test and report gene-environment correlation, and main genetic and interaction effects for all scales and statistical models considered (useful reference33)
Quality control of summary statistics	Follow standard recommendations for study-level and meta-analysis summary statistics quality control and consider additional practices when needed for each specific scenario (useful reference17)
Reporting and sharing G × E research
Reporting	Follow STREGA guidelines for genetic association studies140
	Follow the PRS-RS guidelines for PGS141
	Follow the GWAS Catalog guidelines for GWAS summary statistics146
	Report on the model parameters and interaction term(s) as well as the effect estimates and statistical significance of genes, environment and G × E; report G × E results even if not significant146
Data sharing	Share summary statistics to publicly available resources13,147

CNV, copy number variation; G × E, gene–environment interactions; GWAS, genome-wide association study; PGS, polygenic score; PRS-RS, Polygenic Risk Score Reporting Standards; STREGA, STrengthening the REporting of Genetic Association Studies.

G × E frameworks are gaining popularity, likely preceding a wave of new approaches especially tailored for the analysis of large-scale sequencing data. Although ${h^2}_{\text{GE}}$ MR and PGS continue to attract attention, the utilization of case-only and family-based designs has declined, giving way to case-control or exposure-stratified GWIS. Along the same line, methodologies focused on non-linear or non-constant effects, longitudinal data, machine learning, Bayesian approaches or vQTLs remain outliers in current research practice.

Nevertheless, this is bound to change as refined methods surface. The development and implementation of analytical frameworks that accommodate G × E plasticity, multiple environment or omic layers could be achieved using existing simulated large-scale genomic data¹²⁵ and large-scale studies of diverse participants incorporating standardized protocols for extensive environmental exposure assessment, such as the National Institutes of Health (NIH) All of Us¹²⁶ and Environmental Influences on Child Health Outcomes (ECHO)¹⁹ research programmes.

Multi-omics as a link between genetics and environment

The integration or combination of multiple omics layers could accelerate the discovery of novel genes or the underlying biological mechanisms of the disease (Fig. 1). Gene expression and DNA methylation patterns are, at least partially, genetically controlled, with differences across populations defined by genetic similarity and/or racial or ethnic identity¹²⁷,¹²⁸. In turn, DNA methylation patterns are associated with a wide range of environmental exposures even from the prenatal stage and can act as a regulator of gene expression¹²⁹. A recent study combined summary statistics for gene-lifestyle interaction associations, DNA methylation and gene expression levels to prioritize relevant loci based on the molecular evidence of these interactions¹³⁰ (Fig. 1C). Another study evaluated changes in allelic gene expression in response to several treatments in pluripotent stem cells from six donors¹³¹. Approximately half of their findings had not been previously reported by large-scale allelic expression analyses not considering interactions, highlighting the potential of constraining genomic variance. The successful application of multi-omics for G × E identification, characterization and prediction requires the development of cost-effective data storage and analysis strategies to address challenges related to high dimensionality, batch effects and temporal heterogeneity on the observation of biological effects across omic layers.

Compounded needs for diversity of genetics and environment for G × E

Knowledge of the G × E architecture of human traits remains limited, partially due to the modest number of G × E studies conducted to date (Fig. 4a), representing 1.14% of the studies and 1.87% of the GWAS Catalog signals (Box 2; see Supplementary information). As is the case in genetics and genomics research at large, another ongoing limitation of current G × E studies is the lack of diversity in genetic studies (Fig. 4b). This limitation is noteworthy as historically under-studied groups may encounter environmental health disparities, such as differential exposure to pollutants or nutritional deficits, that could influence the gradient of G × E effect sizes estimates at causal loci across populations¹³². Moreover, differences in allele frequency distributions and linkage disequilibrium patterns across populations could also affect the interaction with the environment, leading to spurious G × E signals if local ancestry is not accounted for¹³³. The presence of shared local environment and possible correlations between genetic ancestry and socio-economic/cultural factors may induce ancestry-environment correlation^134,135, which needs to be considered when evaluating differences across populations and environments.

Fig. 4 |. GWIS characteristics and findings over time, 2008–2022, as curated by the NHGRI-EBI GWAS Catalog.

a, Maximum and median number of participants of the genome-wide interaction study (GWIS) discovery sample, cumulative number of gene-environment interactions (G × E) signals and GWIS publications with and without full summary statistics per year. b, GWIS individual sample size across the discovery stage per study accession and ‘broad’ ancestry category per year. The barplot summarizes the data between 2008 and 2022. c, Outcome distribution per ‘broad’ ancestry category as annotated by the GWAS Catalog based on the number of publications over time. d, Exposure distribution per ‘broad’ ancestry category as annotated by the GWAS Catalog based on the number of publications over time. GWAS, genome-wide association study; NHGRI, National Human Genome Research Institute.

Box 2. The current landscape of G × E.

As of March 2023, the National Human Genome Research Institute (NHGRI)-EBI GWAS Catalog¹⁴⁷ encompasses 228 gene-environment interactions (G × E) publications spanning 806 unique study accessions — trait-specific analyses conducted within each publication — across a total of 649 unique outcome-exposure pairs and 9,240 G × E signals. The number of summary statistics publicly available has not kept pace with the rising sample size and identified G × E signals (Fig. 4a), limiting their downstream use. Although the largest meta-analysis across the discovery and replication phases included 610,410 diverse participants (Box 1), most studies were much smaller (50.9% under the median, ~13,600).

As observed for genome-wide association studies (GWAS)¹³⁴, the majority contribution to G × E studies derives from individuals of European ancestries in terms of sample size (87.3% and 65.4% in the discovery and replication phases, respectively) and assessed outcomes (Fig. 4b–d). The most frequently investigated outcomes across ancestry groups are lipid/lipoprotein (n=27) and cardiovascular measurements (n=24), as well as biological processes (n=18), excluding general broad ‘other’ categories (Fig. 4c). The most frequently considered exposures include alcohol consumption (n=40), smoking exposure (n=29) and sex (n=22) (Fig. 4d).

Despite the absence of a harmonized database for polygenic G × E, there is a discernible surge in the number of publications on polygenic score-environment interactions (PGS×E) (n=294, from 2008 until 2022) compared with genome-wide interaction studies (GWIS), with a predominant focus on psychiatric (22.7%), neurological (13.4%) and psychological (12.0%) traits (see the figure, parts a,b). Notably, a considerable proportion of publications are authored by researchers affiliated to institutions in the United States and England (72.9%) and funded by the US Department of Health and Human Services (HHS) and UK Research and Innovation (UKRI) institutions (see the figure, parts c,d). Although these trends are consistent with large-scale genomic research in general, this emphasizes a need for better representation across outcomes, environments and geography to fully explore the heterogeneity of genetic effects by diverse environmental contexts.

Number of GWIS and PGS × E publications published per year between 2008 and 2022 (part a). Polar barplots for cumulative percentages per study type by research areas (part b), country of affiliation (part c) or funding source (part d). Top ten items are shown. CIHR, Canadian Institutes of Health Research; EU, European Union; MRC, Medical Research Council; NCI, National Cancer Institute; NIA, National Institute on Aging; NIH, National Institutes of Health; NSFC, National Natural Science Foundation of China.

Beyond the genetic ancestry-environment correlation, appropriate strategies are needed to address environmental geographic heterogeneity. Foremost, excluding such diversity may restrain the reach of public health initiatives, especially for collectives affected by environmental geographic inequalities. Conversely, biased findings may arise if inadequately considered. Homogeneous subgroup stratification could improve signal detection by constraining either the environmental or the genomic variance¹³¹, as long as the phenotypic variance is not diminished to the point that it compromises the research question. Although there is no consensus on the strategy for subgroup stratification, exploratory and sensitivity analysis could discern the effects of sample assignment by self-reported ethnicity or major ancestral groups, genomic clusters (for example, estimated identity-by-descent haplotype groups or global ancestry) or environmental/demographic clusters (for example, age groups or school district).

Further international collaborations and initiatives diversifying the ancestry, geography and socio-economic backgrounds of study participants and research teams hold the potential to foster a more robust understanding of the full spectrum of genetic effects across local contexts which require nuanced understanding that would not be achieved without local expertise. This strategy could help mitigate the Eurocentric biases in G × E studies (Fig. 4b–d and Box 2) and limited regional diversity in the research workforce, ultimately shaping the scientific questions posed and funded.

Adherence to FAIR data principles for G × E

The rise in biobank and consortia initiatives has solidified summary statistics as the standard for data sharing, considering the stringent requirements for individual-level data in terms of data storage memory and anonymity protection. Although some of the discussed G × E frameworks use summary statistics to advance our understanding of trait and disease mechanisms, few publications openly shared these statistics through the GWAS Catalog¹³⁶ (Fig. 4a). The PGS Catalog¹³⁷, although currently lacking PGS × E, has the potential for their incorporation. Promoting open and consistent data sharing of G × E findings, under the FAIR (Findable, Accessible, Interoperable and Reusable) data principles¹³⁸, should be a priority for journals and researchers. These practices enhance not only reproducibility, portability and transparency assessments but also dissemination of research results¹³⁹ and the development of new analytical frameworks.

Given the wide breadth of analytical G × E frameworks and the presence of standardized reporting frameworks for some, such as STrengthening the REporting of Genetic Association Studies (STREGA)¹⁴⁰ for GWAS and Polygenic Risk Score Reporting Standards (PRS-RS)¹⁴¹ for the PGS, extensions should be developed to account for current approaches to G × E analyses. In this regard, there is a pressing need for guidelines in constructing, evaluating and reporting PGS × E, with best practices involving untangling potential confounding, interaction and correlation factors. Despite the seemingly modest contribution of G × E to human health^38–42, the assessment of variation in heritability across environmental contexts could provide insights into the potential clinical utility of PGS developed considering specific high-risk exposure strata (Box 1).

Conclusions

The aetiology of human diseases is multifaceted and intricate, often involving the interplay of various genetic and environmental factors. The exploration of G × E in this context faces substantial methodological and conceptual challenges. Despite the variety of G × E statistical frameworks for multiple scenarios, few methods accommodate G × E plasticity, and large-scale studies of these questions remain scarce. As the number of G × E studies on human traits and diseases continues to rise, it is crucial to emphasize the need for diversity and inclusion, standardized reporting protocols and data sharing. These practices will foster transparency and facilitate the assessment of reproducibility in G × E research. Lastly, there is a need for frameworks that consider the spectrum of genetic variation types as the field has increasingly moved towards their inclusion. Although G × E may not account for the bulk of phenotypic variability for complex human traits and diseases, the identification of G × E may increase our understanding of the underlying biological mechanisms and inform environmental health decisions and risk assessment for specific genetic subgroups, providing a much-needed additional link between the success of traditional trait-mapping approaches and their translation for precision medicine and public health.

Supplementary Material

The online version contains supplementary material available at https://doi.org/10.1038/s41576-024-00731-z.

Glossary

Biological or functional interaction

The scenario whereby the exposures physically interact to produce an outcome

Broad-sense heritability

(H²). The proportion of phenotypic variation attributed to genetic variation, including additive, dominance, epistasis and gene-environment interactions (G × E)

G × E correlation

(r_GE). A situation when there are genetic differences in exposure to specific environments. rGEcan be classified as passive, active or evocative according to its source

G × E heritability $\left({h^2}_{\text{GE}}\right)$

The proportion of the phenotypic variability attributable to G × E

Genome-wide interaction studies

(GWIS). Analytical approaches to investigate the modifying effect of an exposure variable on a genetic association

Heteroscedasticity

A statistical phenomenon where the variability of the residuals (errors) in a regression model varies across different levels of the predictor variables

Imputation

The statistical process of predicting missing genetic data based on observed data, typically using reference panels of known genetic variants

Instrumental variables

Variables associated with an exposure that is not associated with the outcome through any other pathway

Mechanistic interactions

Situations when there are individuals for whom the outcome would occur if both of two exposures were present but not if one or both of the exposures were absent

Mendelian randomization

(MR). An analytical method that uses ‘instrumental variables’ for the predictor to determine the causal influences of a predictor on an outcome. MR is based on three key assumptions: instrumental variables are associated with the exposure; instrumental variables are not associated with any confounder; and instrumental variables do not exert effects on the phenotype not mediated by the exposure (also known as ‘horizontal pleiotropy’)

Multiple testing

Repeated statistical testing of hypotheses within a single study, increasing the likelihood of false positives. Adjustments such as Bonferroni correction help control the family-wise error rate

Polygenic score

(PGS). A single value that aggregates the effect of variants across an individual’s germ-line genome to quantify the genetic predisposition to a trait. Usually calculated as a weighted sum of several trait-associated alleles (dosage), where the per-allele effect sizes (β_i) are extracted from the genome-wide association study (GWAS) summary statistics of independent studies

Set-based analyses

Analytical methods that aggregate G × E signals within a gene or genomic region

Statistical interactions

Situations when the effect of a risk factor on an outcome varies across strata of another factor. An interaction on an additive scale occurs when the combined effect of the risk factors differs with regards to the sum of the individual effects, whereas an interaction on a multiplicative scales occurs when the combined effect of the risk factors differs with regards to the product of the individual effects

Variance quantitative trait locus

(vQTL). Genetic variants associated with the variance of a continuous trait