fastGxE maps gene–environment interactions across UK Biobank traits

A scalable statistical method called fastGxE enabled a genome-wide search for gene–environment interactions across 32 physical traits and 67 blood biomarkers in UK Biobank participants, identifying seven loci associated with physical traits and 18 associated with biomarkers. The study, published in Nature Genetics in 2026, reported 16 loci that had not previously been associated with these interactions in the analyzed traits. The results show that measured environmental exposures can modify the observed influence of individual genetic variants, while the authors emphasize that the method remains limited to quantitative traits and one genetic ancestry at a time.

Most conventional genome-wide association studies estimate the average effect of a genetic variant across a population. Such main-effect analyses can miss variants whose influence becomes stronger, weaker or reversed in particular environments. fastGxE was designed to test whether a SNP interacts with one or more measured exposures while accounting for several features of large biobank data. Its model separates the SNP’s main effect, the main effects of environmental factors and the SNP-by-environment interaction terms. It also models polygenic effects, widespread polygenic interaction effects and variation in noise across environments, allowing it to test many exposures without assuming that their effects are unrelated.

Technically, fastGxE examines one SNP at a time. For each possible sparse configuration in which the SNP interacts with a single environmental factor, it performs a score test. It also applies a variance-component test designed for the possibility that the SNP has small interaction effects with many factors. The resulting P values are combined with the Cauchy combination rule, which can combine potentially correlated tests without a loss of type I error control. This mixture of sparse and many-small-effects assumptions is important because the number of true interacting environments is unknown in advance, and a single global test can lose power as the number of measured factors grows.

The method also includes mmSuSiE, a mixed-model extension of the SuSiE algorithm, for evaluating which environmental factors are associated with a detected interaction. For each significant SNP, mmSuSiE fits the candidate SNP-by-environment terms while retaining polygenic main effects, polygenic interaction effects and heterogeneous environmental effects. It reports posterior inclusion probabilities for the environmental factors, which are used to identify factors contributing to the interaction. A separate data-splitting analysis estimates interaction effects in one half of a sample, assigns participants in the other half to five groups using an aggregated interaction environment score and tests whether the SNP’s main effect differs among those groups. These downstream analyses were designed to characterize and corroborate signals identified in the genome-wide scan.

In computational benchmarks, fastGxE expanded the practical scale of these analyses. For 300,000 individuals, 7,833,003 SNPs and 64 environmental factors, the method completed the scan in 6.3 hours using 16 CPU cores. That was 32.98 times faster than the single-environment method fastGWA-GE and 126.49 times faster than the multi-environment method StructLMM in the study’s comparison. A block-wise parallelization strategy also improved performance when analyses were distributed across multiple nodes. The speed advantage persisted as the number of cores, environmental factors and participants increased, indicating that the method was designed for the scale of biobank-sized datasets rather than only small validation samples.

Simulations supported the method’s statistical performance but also defined its limits. Across null settings, fastGxE maintained well-calibrated P values while accounting for SNP heritability, polygenic interaction heritability and environment-dependent noise heterogeneity of 0%, 15% or 25%. It generally had greater statistical power than fastGWA-GE and StructLMM as interaction effects became smaller, the number of active environments increased or some true interacting environments were omitted from analysis. The method also remained calibrated in simulations incorporating 96 genome-wide annotations, and it produced unbiased estimates of overall gene–environment interaction heritability. However, inflation appeared in a restricted set of proxy scenarios involving a rare causal variant, a relatively strong marginal genetic effect and moderate linkage disequilibrium between the causal and tested SNPs.

The physical-trait analysis included 214,580 to 324,811 participants and 42 environmental factors drawn from demographic, lifestyle, dietary and social domains. Estimated SNP heritability ranged from 16.67% to 67.35%, with a mean of 28.74%. Polygenic gene–environment interaction heritability was smaller, ranging from 0.96% to 10.48% and averaging 6.20%, while environment-dependent noise heterogeneity ranged from 1.61% to 25.24%. The researchers found seven genomic loci containing 23 genome-wide significant interactions across 20 physical traits. Three loci, including FTO, CHRNA5-A3-B4 and WNT16, extended previous evidence, while FBXO32, LYPLAL1-AS1, COBLL1 and RSPO3 were reported as novel discoveries.

Environmental influences differed among the physical-trait signals. mmSuSiE identified an average of 2.87 environmental factors per significant SNP and 66 SNP–environment pairs across the 20 traits. Age, alcohol-intake frequency and stair-climbing frequency appeared frequently among the leading factors. In the FTO locus, interactions involving 15 traits included age, alcohol consumption, stair climbing and several dietary or sleep-related variables. At CHRNA5-A3-B4, the lead SNP rs8042849 was associated with the forced expiratory volume to forced vital capacity ratio; water intake attenuated its minor-allele effect, whereas smoking status and television-watching time increased it. The FBXO32 signal for pulse pressure was attributed primarily to age. These are statistical associations between measured exposures and variant effects, not evidence that changing any exposure would produce a clinical benefit.

The biomarker analysis included 178,899 to 312,887 participants per trait. Mean SNP heritability across the 67 biomarkers was 29.14%, compared with 2.67% for polygenic gene–environment interaction heritability. fastGxE detected 18 loci containing 38 significant interactions across 19 biomarkers. Six loci had previous interaction evidence, while 12 were described as novel. mmSuSiE identified an average of 3.55 contributing environments per significant SNP and 135 SNP–environment pairs overall. Age, processed-meat consumption and bread intake were among the most frequent factors. The APOE locus produced interactions with total cholesterol, low-density lipoprotein and apolipoprotein B, with age, bread intake, smoking and television-watching time implicated as drivers. The analysis did not replicate five previously reported loci, although the estimated interaction effects at four of those loci were directionally consistent with earlier reports; one previously significant SNP was excluded after failing a Hardy–Weinberg equilibrium filter.

The researchers used the data-splitting procedure to assess the detected interactions. Twenty-one of 23 physical-trait signals and 36 of 38 biomarker signals showed differences among the five score groups at thresholds corrected for the number of associations tested. The analysis provides additional statistical support for heterogeneity in SNP effects across environmental-score groups, but it does not establish that the associated SNPs are causal or that the measured environments cause the observed changes. The authors also note that a significant SNP may tag a causal variant through linkage disequilibrium, so the associations do not identify the precise causal genetic variant by themselves.

Several limitations constrain the conclusions. fastGxE was developed primarily for quantitative phenotypes; binary case–control data can be approximated by treating the outcome as continuous, but a principled discrete-outcome extension remains to be developed. The method does not yet perform dedicated fine-mapping of interaction effects, and it analyzes one genetic ancestry at a time, limiting the borrowing of information across ancestries. No mental-health-related interactions remained significant after multiple-testing correction, so the results should not be generalized to other health domains without further evidence. The authors estimated that UK Biobank-sized samples can reliably detect interactions explaining more than 0.04% of phenotypic variance and can retain reasonable power above 0.02%. Larger effects may be detected with roughly 100,000 participants and moderate effects with about 300,000, whereas small effects can require approximately 800,000 participants and very small effects more than one million. The study therefore demonstrates both the scalability of fastGxE and the statistical limits of current cohort sizes for detecting small interaction effects.

Subject of Research: Scalable genome-wide detection of gene–environment interactions in UK Biobank physical traits and blood biomarkers

Article Title: Genome-wide detection of genotype–environment interactions at biobank scale with fastGxE

Article References: Ning, C., & Zhou, X. (2026). Genome-wide detection of genotype–environment interactions at biobank scale with fastGxE. Nature Genetics. https://doi.org/10.1038/s41588-026-02809-2

Image Credits: AI Generated

DOI: 10.1038/s41588-026-02809-2

Keywords: fastGxE, gene–environment interaction, UK Biobank, genome-wide association, physical traits, blood biomarkers, statistical genomics, polygenic interactions, biobank-scale analysis, genetic ancestry, Genome-wide, detection