Gene-Environment Interaction — GxE, GWIS, MR, Epigenetic Mediation

Short version. Gene-environment interaction is the regression interaction term β_GxE on top of the additive genetic and environmental main effects. It is statistically hard to detect at scale because well-powered GxE designs need much larger samples than additive-effect designs, and because environmental exposures are typically poorly measured. Classical examples (PKU, G6PD, FTO, CYP2A6) are unambiguous because the exposure is sharp and the genetic effect is large. Mendelian randomisation reframes the question by treating genetic variants as instrumental variables for environmental exposures. Epigenetic mediation — DNA methylation as a G-E interface — is the mechanism evidenced by the Dutch Hunger Winter cohort, where periconceptional famine left a persistent IGF2 hypomethylation signal six decades later. None of this is clinical advice; framing here is research and education.

The GxE definition

For an outcome Y, a genetic variable G (a single SNP, a polygenic score, or a haplotype), and an environmental variable E (a continuous or categorical exposure), the conventional regression model is:

Y = μ + β_G·G + β_E·E + β_GxE·(G·E) + ε

where β_G and β_E are the additive main effects and β_GxE is the interaction term. The null hypothesis of no GxE is β_GxE = 0; rejection of the null is the statistical evidence of interaction. In practice the interaction term is interpreted on the scale of the analysis — on the linear scale for continuous outcomes, on the log-odds scale for binary outcomes — and reanalysis on alternative scales sometimes resolves apparent interactions away.

Two distinctions in the GxE literature are worth keeping straight. First, statistical GxE is scale-dependent: an interaction in a logistic-regression model may not appear in a linear-probability model and vice versa. Second, biological GxE refers to mechanism: the same genotype produces different phenotypes in different environments because some pathway depends on both. Statistical GxE is necessary but not sufficient evidence for biological GxE. Studies should declare the scale of analysis and treat scale-dependent interactions with appropriate caution.

Statistical power: why GxE is hard

The minimum sample size to detect a GxE effect of given magnitude is several times larger than to detect an equivalent main effect, because the interaction term has a smaller signal-to-variance ratio in any reasonable allele-frequency / exposure-distribution combination. Hunter 2005 (Nat Rev Genet 6:287) is the canonical methodological treatment for human GxE studies, setting out three issues that drive the difficulty:

Exposure measurement error. Environmental exposures are usually self-reported, retrospectively assessed, or coarse-categorical (smoker / non-smoker), each of which biases the GxE estimate toward the null.
Multiplicative-effect distributions. Where the gene and environment have effects of different orders of magnitude, the interaction term is dominated by whichever has lower precision — usually the environment.
Multiple testing. Genome-wide GxE search across millions of variants and one or more exposures invites Bonferroni penalties orders of magnitude harsher than additive GWAS.

A consequence is that almost all genome-wide GxE findings reported with confidence in the contemporary literature have come from very large consortium analyses, careful exposure phenotyping, and pre-registered hypothesis sets. Mostafavi et al. 2020 (PLoS Biol 18:e3000812) discuss the broader portability and predictive-stability problems of polygenic prediction across environments and time, of which GxE is a part.

Classical examples

The unambiguous GxE examples in human genetics are those where the environmental exposure is sharp and the genetic effect is large. Six are routinely cited.

Phenylketonuria (PKU) and dietary phenylalanine. Homozygous loss-of-function variants in PAH produce severe intellectual disability on an unrestricted diet and largely normal cognitive outcomes on a low-phenylalanine diet introduced in infancy. The same genotype, two outcomes; the difference is environmental. PKU is the canonical illustration of GxE in introductory teaching and underwrites the case for newborn screening of inborn errors of metabolism.
Glucose-6-phosphate dehydrogenase deficiency (G6PD) and oxidant exposures. X-linked variants reducing G6PD activity precipitate haemolytic anaemia on exposure to fava beans, certain antimalarial drugs, sulphonamides, and other oxidant agents; carriers are largely asymptomatic in the absence of the trigger. Exposure-conditional phenotype.
Alcohol dehydrogenase / aldehyde dehydrogenase variants and alcohol. The ALDH2*2 variant common in East Asian populations produces flushing, tachycardia, and toxic acetaldehyde accumulation after even modest alcohol intake; in non-drinkers the phenotype is silent.
CYP2A6 variants and tobacco smoking. CYP2A6 is the principal nicotine-metabolising enzyme; reduced-activity variants alter nicotine clearance, smoking topography, and smoking-related disease risk — conditional on smoking, with no effect in non-smokers.
FTO and physical activity. The FTO rs9939609 obesity-associated variant has a genetic effect on BMI substantially attenuated in physically active adults compared with sedentary adults. Multiple meta-analyses have replicated the interaction.
Skin-pigmentation variants and ultraviolet exposure. Variants at MC1R, OCA2, SLC45A2 and others affect melanoma risk in interaction with cumulative UV exposure; the effect of any one variant is meaningfully different in low-UV vs high-UV environments.

None of these examples is interpreted on this site as a basis for individual clinical decisions; the framing is research and education.

Genome-wide interaction studies (GWIS)

A GWIS is the GxE analogue of a GWAS: scan the genome for variants whose marginal effect on Y depends on the environmental exposure E. The naive design tests β_GxE for every genotyped variant and applies a Bonferroni or FDR correction. Power is poor at any reasonable sample size, and the literature has converged on two design improvements.

Two-stage filtering. First, screen variants by some marker correlated with interaction signal (e.g. heteroskedasticity in Y across genotype groups) to reduce the multiple-testing penalty in the formal GxE test. Several variants of two-stage screening are in use; all carry careful caveats about preserving the type-I error rate.
Case-only design. Among cases of a binary disease, the dependence of E on G provides a test of GxE under the assumption that G and E are independent in the source population. This sidesteps the recruiting of unaffected controls but inherits the strong G-E independence assumption, which can fail for traits whose underlying biology drives behavioural exposure choices (e.g. high-genetic-risk smokers selecting in or out of smoking).

Software packages such as PLINK and the EMMAX/MMM family implement formal GxE tests; specialised pipelines such as GxEScan and CGEN are available for the two-stage and case-only designs.

Mendelian randomisation: instrumental variables for environmental causality

Mendelian randomisation reframes a related question: not whether gene and environment interact, but whether an environmental exposure causally affects an outcome. Davey Smith & Hemani 2014 (Hum Mol Genet 23:R89; PMID 25064373) is a canonical methodological review. The technique uses one or more genetic variants as instrumental variables for the environmental exposure. The variant must satisfy three core assumptions: it is associated with the exposure (relevance), it is independent of unmeasured confounders of the exposure-outcome relationship (independence), and it acts on the outcome only through the exposure (exclusion restriction).

If the assumptions hold, the ratio of the variant-outcome effect to the variant-exposure effect estimates the causal effect of the exposure on the outcome — the Wald estimator. With multiple instruments, two-stage least squares or inverse-variance-weighted meta-analysis aggregates the per-instrument estimates. MR-Egger regression explicitly tests and partially corrects for horizontal pleiotropy, in which a variant affects the outcome through pathways other than the exposure. Sensitivity analyses such as MR-PRESSO, weighted median, and weighted mode estimators provide robustness checks against assumption violations.

MR has been deployed to interrogate the causal status of many established epidemiological associations. C-reactive protein and coronary heart disease, vitamin D and various outcomes, alcohol intake and cardiovascular disease, and BMI and education attainment are among the well-known reassessments — some of which dissolved under MR scrutiny, others of which survived. MR results are research and education content on this site, not clinical guidance.

Epigenetic mediation and the Dutch Hunger Winter

An influential mechanistic frame for GxE is that environmental exposures alter epigenetic state — chiefly DNA methylation — and that the resulting epigenetic state mediates downstream phenotype. The strongest human-cohort evidence comes from Heijmans et al. 2008 (PNAS 105:17046; PMID 18955703), the Dutch Hunger Winter study. The cohort comprises adults who were exposed in utero to the Dutch famine of 1944–1945, in which calorie intake fell to ~500–1000 kcal per day across western Netherlands for several months under wartime blockade.

Heijmans and colleagues compared DNA methylation at the imprinted IGF2 locus in cohort members exposed periconceptionally with that of their unexposed same-sex siblings, six decades after the famine. Periconceptionally exposed individuals showed significantly lower IGF2 methylation, persisting into late adulthood. Subsequent analyses extended the methylation differences to other loci and connected them, in some cases, to adult metabolic phenotype. The Dutch Hunger Winter is now the canonical illustration of the persistence of environmentally induced epigenetic marks across decades and a central reference in the developmental-origins-of-health-and-disease (DOHaD) literature.

Subsequent prenatal-stress cohorts (Project Ice Storm in Quebec, the Chinese Famine birth-cohort studies, several pollution-exposure cohorts) have replicated the structure of the Dutch Hunger Winter findings with varying signal magnitudes. The collective body of work has motivated the methylome-wide association study (EWAS) as a methodology in its own right.

Epigenome-wide association studies (EWAS)

An EWAS is the methylation analogue of a GWAS: across hundreds of thousands of CpG sites measured by Illumina Infinium arrays (the 27K, 450K, EPIC, and EPIC v2.0 platforms), test the association of methylation level at each site with a phenotype, exposure, or trait. The EWAS Catalog and the EWAS Atlas curate replicated findings across cohorts.

Three problems shape the EWAS literature and should be borne in mind by readers. First, tissue specificity: methylation patterns vary across tissues, and most human EWAS use peripheral blood as a proxy, which may or may not reflect the tissue of relevance for the phenotype. Second, cell-type heterogeneity: methylation differences between cases and controls can reflect altered blood-cell composition rather than altered methylation within a fixed cell type, and reference-based or reference-free deconvolution is now standard. Third, causality direction: methylation differences may cause the phenotype, be caused by it, or share a common cause; EWAS is observational, and causal inference requires further work (Mendelian randomisation, longitudinal sampling, in vitro perturbation).

None of this content is clinical guidance. EWAS is a research tool; methylation findings on this site and in the help catalogue are framed as research and education.

GxE in the broader polygenic framework

The pure additive polygenic decomposition assumed in the polygenic models page treats V_G and V_E as independent and additive. Substantial GxE means that this decomposition is misspecified: an interaction term V_GxE is needed, and heritability estimates derived under the additive assumption may be biased depending on the magnitude and structure of the interaction. In practice, for most complex traits, V_GxE appears to be small relative to V_A and V_E, but this is a quantitative claim that depends on the trait, the environment, and the population studied.

For polygenic risk scores, GxE means that an individual's PRS-implied risk depends on environment as well as genotype. A PRS for type 2 diabetes computed in a sedentary population may overstate risk in a physically active subgroup (the FTO illustration generalises). PRS calibration across environmental subgroups, like PRS calibration across ancestries, is a research-grade question and not a settled one. Treating PRS output as a fixed personal attribute independent of environment is the wrong reading; treating it as a population summary that depends on the validation cohort's environmental distribution is the right one.

Gene-environment interaction: study designs and epigenetic mediation