Main

There has been great success in human genetics—particularly genome-wide association studies (GWAS)—at revealing disease pathophysiology and complex traits biology4. Genetic association mapping on multi-omics layers has covered proteomics5, metabolomics6 and single-cell RNA sequencing (scRNA-seq)7, providing granular insights into trait-associated genetic loci. However, such efforts focus on fixed genetic effects (more precisely, marginal effects), oversimplifying the intrinsic complexity of trait biology1. Essentially, human phenotypes show dramatic changes in response to multifactorial environmental exposures, including sex, senescence and lifestyle. Inter-individual heterogeneity in responses to environments has been shaped by genetic adaptation8,9 and affects present disease risks2 and drug efficacy10. Genetically, this phenotypic plasticity manifests as changes in genetic effect sizes across environmental factors (or, equivalently, changes in environmental effects across genotypes), namely, gene–environment (G×E) interactions3 (Fig. 1a). G×E interactions capture dynamic changes in genetic effects, unveiling the genetic regulation of phenotypic plasticity. In some traits, G×E interaction studies have begun to explain phenotypic variation not captured by marginal effects (that is, missing heritability) and disparities in polygenic risk prediction11. Therefore, identifying G×E interactions may contribute to mitigating health disparities and implementing personalized medicine precisely12.

Fig. 1: A cross-population atlas of G×E interactions.

a, A schematic of gene–environment (G×E) interactions. b, Overview of the cohorts. EUR, European; AFR, African; AMR, American; EAS East Asian. c, Circular plot of genome-wide significant G×E interactions in UKB, estimated using two-sided linear regression for quantitative traits and Firth logistic regression for dichotomous traits. Dot sizes represent replication P values in UKB2 and BBJ2. Pleiotropic loci are indicated by lines, with the nearest genes to the lead variants. d, Numbers of genome-wide significant loci per trait in UKB. Bar colours represent trait categories; the colour intensity indicates the types of pleiotropy. e,f, As in d (e) and c (f), for BBJ (see Supplementary Table 1 for the sample sizes used in c–f). AST, aspartate aminotransferase; ALT, alanine aminotransferase; BC, red blood cell; BS, blood sugar; CAD, coronary artery disease; CRP, C-reactive protein; Eosino, eosinophils; Hb, haemoglobin; Ht, haematocrit; Lymph, lymphocytes; MCH, mean corpuscular haemoglobin; MCHC, mean corpuscular haemoglobin concentration; MCV, mean corpuscular volume; P, phosphate (UKB)/inorganic phosphorus (BBJ); Plt, platelets; PP, pulse pressure; sCr, serum creatinine; T2D, type 2 diabetes; T.Bil, total bilirubin; TC, total cholesterol; TG, triglycerides; TP, total protein; UA, uric acid.

Nevertheless, after decades of effort3, there is a limited number of established G×E interactions in humans13, and their biological interpretation remains underestablished14. Past studies have suffered from low replication rates15,16 due to low statistical power17, heavy multiple-testing burdens18, arbitrary filtering of genetic variants15 and, in some cases, imprecise statistical testing19. G×E interactions have been studied at scale only for limited traits and environments, primarily in European populations20. Therefore, a global overview of G×E interactions across phenotypes, environments and populations remains unknown.

Here, using the recent advent of population-scale biobanks21 and computationally efficient methods22, we conducted parallel genome-wide G×E interaction studies using UK Biobank (UKB) and Biobank Japan (BBJ) to provide a cross-population atlas of G×E interactions. We validated the identified G×E interactions in four independent cohorts with diverse populations, annotated the environmental contributors and assessed their impacts on heritability, polygenic prediction accuracy and responsible cell types. Multi-omics G×E analyses provided molecular insights into clinical G×E interactions. These multi-resolution analyses demonstrated that G×E interactions have pivotal roles in regulating dynamic phenotypic plasticity, informing personalized phenotype prediction and drug development.

G×E interactions in individual biobanks

To reliably detect G×E interactions, we divided UKB and BBJ into discovery and replication cohorts (UKB1, Nmax = 273,453; UKB2, Nmax = 38,149; BBJ1, Nmax = 166,757; BBJ2, Nmax = 65,373) (Fig. 1b and Supplementary Table 1). Targeting 38 biomarkers and 9 diseases, which spanned 10 categories (anthropometric, metabolic, proteins, kidney-related, electrolytes, liver-related, inflammatory, haematological, blood pressure and diseases), G×E interactions were tested for nine environmental factors individually and jointly, with P values aggregated per variant on the Cauchy distribution23 to assess genome-wide significance (Extended Data Fig. 1 and Supplementary Tables 2 and 3). The environmental factors included age, sex, ever-drinking, ever-smoking and current-smoking, and four clusters for diet and physical activity derived from questionnaire data (Extended Data Fig. 2 and Supplementary Table 4).

In UKB1, we identified 64 genome-wide significant G×E interactions at 45 loci spanning all trait categories (PG×E < 5.0 × 10−8), with 31 interactions at 23 loci remaining after Bonferroni correction (PG×E < 5.3 × 10−10), indicating that G×E interactions were widespread in human complex traits (Fig. 1c,d and Supplementary Tables 5 and 6). These included known interactions, G×Current-smoking at the HYKK locus for body mass index (BMI)24, G×Age at the UMOD locus for estimated glomerular filtration rate (eGFR)25, and G×E at the FTO locus for BMI driven by multiple environments such as physical activity, diet, age, drinking and smoking26, empirically validating our results. The remaining interactions—based on our curation of GWAS Catalog27—were not reported at PG×E < 5.0 × 10−8 (Supplementary Note 2 and Supplementary Table 7). In total, 16 loci overlapped with recent UKB G×E reports using different protocols13,28 (Supplementary Note 3 and Supplementary Table 8). We observed pleiotropy at 13 loci (10 intra- and 3 inter-categorical), with 2 inter-categorical loci showing distinct significant variants across trait categories (Extended Data Fig. 3), suggesting trait-category specificity of G×E pleiotropy in contrast to the broader pleiotropy of marginal effects29.

In BBJ1, 36 significant G×E interactions were detected across 15 loci (26 across 8 loci after Bonferroni correction) (Fig. 1e,f and Supplementary Table 6). These included the well-established locus in the European population, that is, the FTO locus for BMI, which we confirmed in the East Asian population (driven by G×Age and G×Ever-drinking). Other loci with PG×E < 5.0 × 10−8 have not been reported in the GWAS Catalog, emphasizing the importance of studying G×E interactions in non-European populations. Notably, 58% (21 out of 36) of G×E interactions were at the ALDH2 locus, which harbours an East-Asian-specific missense variant (rs671) with a strong dominant effect on alcohol metabolism30, consistent with its high pleiotropy for GWAS29.

Inflation was minimal in both cohorts (Supplementary Table 5), and the results were robust to phenotype normalization (Supplementary Table 9). A stepwise variable selection approach revealed that a mean of 2.4 environments contributed to G×E interactions (range = 1–7; Supplementary Table 6), supporting our approach of testing G×E interactions both individually and jointly across environments. In 91% (10 out of 11) of the intra-categorical pleiotropic loci, at least one shared environment contributed to all traits, suggesting that the combinations of trait categories and environments were major determinants of G×E pleiotropy.

Replication within populations

Of the 64 G×E interactions in UKB1, 23 were nominally replicated in UKB2 (PG×E < 0.05), and 6 remained significant after Bonferroni correction (Fig. 1c and Supplementary Table 6). In BBJ, 28 out of 36 G×E interactions were nominally replicated, and 19 were significant in BBJ2 (Fig. 1f and Supplementary Table 6). These included a G×Ever-drinking interaction at the ALDH2 locus for haemoglobin in BBJ1 (PG×E = 2.2 × 10−15; Pmarginal = 1.7 × 10−3), replicated in BBJ2 (PG×E = 2.9 × 10−9; Pmarginal = 8.9 × 10−3), highlighting context-specific effects that would be missed by marginal genetic tests. The same locus also showed a significantly replicated G×Ever-drinking interaction for type 2 diabetes (PG×E = 4.8 × 10−16 in BBJ1 and 1.2 × 10−6 in BBJ2). This interaction remained significant after adjusting for haemoglobin (PG×E = 8.1 × 10−15 in BBJ1), suggesting minimal mediation by haemoglobin. For replication, we stringently required consistent environments and effect directions; failure in either was deemed non-replicated, regardless of P values. Replication rates were comparable with those in GWAS, supporting the robustness of our findings (Supplementary Note 4).

We further tested replication in independent cohorts: the European population in All of Us (Nmax = 208,700) and the East Asian population in two Japanese cohorts: (1) the Japan Multi-Institutional Collaborative Cohort–Hospital-based Epidemiologic Research Program at Aichi Cancer Center (J-MICC/HERPACC) (Nmax = 70,909), and (2) the Japan Public Health Center-Based Prospective Study (JPHC) (Nmax = 10,904) (Supplementary Table 10). Despite differences in lifestyle questionnaires, dietary clusters (for example, ‘Japanese cuisine (Washoku)’ in all Japanese cohorts, and ‘meat and cheese’ in UKB1 and JPHC) were consistently recovered across cohorts, supporting the robustness of our clustering approach (Extended Data Fig. 4). Bonferroni replication rates were 27% in All of Us (17 out of 64 trait–locus pairs) and 56% in J-MICC/HERPACC (20 out of 36; Extended Data Fig. 5a,b and Supplementary Table 11). Notably, the pleiotropic G×E interactions at the ALDH2 locus were replicated at 81% (17 out of 21) in J-MICC/HERPACC, with six also replicated in JPHC despite its smaller sample size.

Collectively, our approach using cross-population biobank resources thoroughly detected and validated G×E interactions across diverse trait categories and environments.

Cross-population consistency

Combining UKB and BBJ results yielded 94 trait–locus pairs across 54 loci. Six loci were shared between biobanks (40% of BBJ1 loci; Fig. 1c,f), often involving essential (‘core’) genes for the target phenotypes. For example, ALPL for alkaline phosphatase (ALP) was commonly driven by G×Sex, GGT1 for γ-glutamyl transpeptidase (GGT) by G×Sex, G×Age and G×Ever-smoking, and UMOD for eGFR by G×Age (Extended Data Fig. 5c–f).

To assess cross-population sharing more broadly, we examined replication in the other population’s discovery cohort. After excluding the ALDH2 locus (rs671) to avoid introducing bias due to its well-known East Asian specificity, 22 out of 73 interactions were nominally (and nine significantly) replicated (PG×E < 6.8 × 10−4). The conservative signal-sharing estimate (Storey’s π1; ref. 31) was 0.41, indicating moderate consistency of G×E interactions across populations. The cross-population-replicated loci included three that were originally from UKB1: G×Age at the APOE locus for total cholesterol; G×Sex at the ABCG2 locus for urate; and G×Sex and G×Physical-activity at the SURF6 locus for ALP (PG×E = 5.7 × 10−4, 3.1 × 10−5 and 1.5 × 10−4 in BBJ1, respectively). These shared G×E interactions suggested that cross-population meta-analyses would be beneficial. Indeed, we detected one additional G×E interaction through a meta-analysis across BBJ and UKB (Supplementary Note 5 and Supplementary Table 12).

Minor allele frequencies of the lead variants tended to be higher in the population where the G×E interactions were detected (Extended Data Fig. 5g). Population specificity (that is, dietary environments) and differing distributions of environments probably also contributed to the population-specific G×E detection (Extended Data Fig. 5h).

We further evaluated replication in the African and American populations in All of Us (Nmax = 70,558 and 66,556, respectively) and the Israeli population in the Human Phenotype Project (HPP; Nmax = 8,645). Three G×E interactions were significantly replicated in the African population—two of which overlapped with UKB1–BBJ1 shared loci (Extended Data Fig. 5a and Supplementary Table 13). In the American population, two G×E interactions for pulse pressure—primarily driven by G×Age—were significantly replicated. Although no G×E interaction reached significance in the Israeli population, possibly due to its small sample size, the top signal aligned with a UKB1–BBJ1 shared interaction. These findings demonstrated both shared and population-specific G×E interactions. Subsampling analyses revealed that detecting G×E interactions required biobank-scale sample sizes, and detected loci were not saturated (Extended Data Fig. 5i), encouraging future global collaboration to thoroughly capture worldwide G×E interactions.

Environments contributing to G×E

Gene–environment interactions can enhance locus interpretation by revealing context-specific genetic associations. In UKB1, diet-related environments contributed to five G×E interactions—all of which are at least nominally replicated. These included the ABCG2 locus, where association with eGFR was specific to non-consumers of ‘meat and cheese’ (PG×E = 1.5 × 10−14; Fig. 2a,b). Raw questionnaire data confirmed that low meat consumption unmasked the genetic effect (Extended Data Fig. 6). As ABCG2 encodes a primarily intestinal urate exporter32 and urate is the end-product of purine metabolism, high purine intake from meat may obscure the genetic effect.

Fig. 2: Representative loci with G×E interactions.

a,b, Urate levels across rs4148155 genotypes (the ABCG2 locus) and ‘meat and cheese’ consumption in UKB1 (a) and UKB2 (b). Dots show 5,000 randomly sampled individual measurements per genotype; lines represent the linear regression lines using all participants. c,d, AST (c) and haemoglobin (d) levels by rs671 genotypes (the ALDH2 locus) and drinking status in BBJ1. Boxplots show the median, quartiles and 1.5× interquartile range (see Supplementary Table 14 for exact P values). e, Schematic of the mechanisms of action for warfarin, natto (fermented soybean) and DOACs. The heart icon is from TogoTV (2016 DBCLS TogoTV, CC-BY-4.0). f, Heatmap of the arrhythmia prevalence in BBJ1. g, As in f, but stratified by arrhythmia subgroups. h, Natto intake before and after warfarin initiation in BBJ atrial fibrillation or flutter patients with ≥1 prothrombin time-international normalized ratio (PT-INR) record per year, considering that regular monitoring of PT-INR is required during warfarin therapy. i, Odds ratios of rs72900155 (the PITX2 locus) for arrhythmia stratified the natto intake frequency. N = 109,642 and 49,116 for natto-consuming and non-consuming participants in BBJ1, and N = 44,954 and 18,155 for those in BBJ2, respectively. Data are presented as estimated values with 95% confidence intervals (Firth logistic regression). j, Heatmap of the atrial fibrillation or flutter prevalence in BBJ2, stratified by warfarin and DOAC medication.

In BBJ1, although ALDH2 primarily affects alcohol metabolism, multiple environments contributed to the pleiotropic G×E interactions at the ALDH2 locus after adjusting for G×Drinking interactions (Supplementary Note 6 and Supplementary Table 14). Stratified analysis in ever- and never-drinkers revealed that 12 of 19 biomarkers showed strong non-additive effects in ever-drinkers (Pnon-additive < 5.0 × 10−8; Fig. 2c and Supplementary Table 15), consistent with the dominant deleterious effect on alcohol metabolism of the lead variant (rs671). Four haematopoietic traits showed purely additive effects in never-drinkers (red blood cells, haemoglobin, haematocrit and white blood cells; Padditive < 5.0 × 10−8 and Pnon-additive > 0.05; Fig. 2d), opposite to the functional role and inheritance pattern of rs671. These effects were all replicated in BBJ2 (Padditive < 0.05/19 = 2.6 × 10−3 and Pnon-additive > 0.05). Given the long-range linkage disequilibrium (~2.44 Mb) with signs of recent selection33 at this locus, other causal variants or genes may underlie these haematopoietic associations. The region harbours haematopoiesis-related genes (for example, SH2B3 and PTPN11), whose roles in common-variant genetics warrant further investigation. To facilitate future research, we applied a deep learning model to prioritize variant–gene pairs with potential regulatory effects (Supplementary Note 7 and Supplementary Table 16). These loci together demonstrated the utility of G×E interactions for gaining biological insights into genetic loci.

A reverse-causal G×E interaction

In BBJ1, the PITX2 locus for arrhythmia showed a G×E interaction primarily driven by natto (fermented soybean) intake (PG×E = 2.8 × 10−12; PG×E = 2.1 × 10−10 when testing G×Natto alone; Extended Data Fig. 7). The lead variant, rs72900155, has been reported to be associated with atrial fibrillation29—a subgroup of arrhythmia. Clinically, warfarin, a long-standing anticoagulant, may link natto and atrial fibrillation. As vitamin K in natto reduces the anticoagulation effect of warfarin, patients on warfarin are advised to avoid it (Fig. 2e). In BBJ1, the arrhythmia prevalence was markedly high in the homozygous carriers of natto non-consumers (Fig. 2f), and this pattern was primarily driven by atrial fibrillation or atrial flutter, for which warfarin was the sole first-line anticoagulant before the launch of direct oral anticoagulants (DOACs) (Fig. 2g). In this subgroup, natto intake declined markedly after warfarin initiation in the same individuals (Fig. 2h), suggesting that this G×E interaction was driven by reverse causality from the disease to the environment.

The marginal effect size of rs72900155 was substantially larger for atrial fibrillation or flutter than for the other subgroups (0.49 (95% CI, 0.46–0.52) versus 0.13 (0.10–0.15)). Owing to this effect size heterogeneity, the overall effect size for arrhythmia would vary with the proportion of atrial fibrillation or flutter patients across natto intake strata, explaining the link between the reverse causality and the G×Natto interaction. Consistently, the G×Natto interaction was not significant in either subgroup when evaluated separately (PG×E = 0.84 for atrial fibrillation or flutter; 2.6 × 10−4 for the other subgroups).

This G×E interaction was not replicated in BBJ2 (PG×E = 0.40; Fig. 2i). Notably, this replication failure might be reasonable as BBJ1 participants were recruited from 2003 to 2008, whereas most BBJ2 participants (84.9%) were recruited from 2013 to 2017, and between these periods, DOACs replaced warfarin in more than half of atrial fibrillation patients in Japan34. As DOACs do not require natto restriction (Fig. 2e), the atrial fibrillation or flutter patients taking DOACs in BBJ2 did not show increased natto avoidance (Fig. 2j).

In summary, we identified the G×E interaction driven by reverse causality. Although machine-learning-based locus interpretation is increasingly investigated35, these results indicate that this technology is not readily applicable to G×E interactions, and careful interpretation by specialists is necessary to disentangle their causal mechanisms.

To evaluate causal directions at other loci, we leveraged repeat biomarker measurements to sort out temporal ordering from environments to phenotypes, and performed time-to-event Cox analyses for disease onset and overall survival (Supplementary Note 8 and Supplementary Tables 17 and 18). We identified a significant G×E interaction for overall survival at the ALDH2 locus driven by interactions with sex, ever-drinking and age (P = 1.7 × 10−11), suggesting potential G×E effects on human lifespan.

Pleiotropic G×E effects on diseases

We conducted phenome-wide G×E interaction analyses to assess pleiotropy on diseases (Supplementary Table 19). In UKB1, we detected one additional G×E interaction at the APOE locus for dyslipidaemia primarily driven by G×Sex (PG×E = 3.8 × 10−7; false discovery rate (FDR) < 0.05), consistent with the G×E interactions in the main analyses for cholesterol biomarkers (total cholesterol, triglycerides and low-density lipoprotein cholesterol (LDL-C); Extended Data Fig. 8 and Supplementary Table 20). In BBJ1, 11 G×E interactions were also significant. The ALDH2 locus exhibited widespread G×E pleiotropy across diseases, including the established G×Drinking interaction for oesophageal cancer36. We also observed a G×Age interaction at the HLA-DQB1 locus for rheumatoid arthritis (PG×E = 1.2 × 10−6), originally detected for asthma in BBJ1 and immune cells in UKB1 (lymphocytes, eosinocytes and white blood cells), suggesting shared G×E effects across immune phenotypes. These results showed that the G×E interactions for clinical biomarkers also affected disease statuses through pleiotropy.

Genome-wide heritability

We estimated G×E heritability to evaluate consistency across populations at the genome-wide level37. Evaluating individual environments, 14 and 12 trait–environment pairs were significant in UKB1 and BBJ1, respectively (FDR < 0.05; Supplementary Table 21), although statistical power was limited by multiple testing burden. When aggregating across environments, G×E heritability was significantly positive for seven traits in UKB1 and 11 traits in BBJ1, including four overlapping traits: height, BMI, high-density lipoprotein cholesterol (HDL-C) and diastolic blood pressure (DBP) (Fig. 3a–c). G×E-to-marginal heritability ratio was much larger for BMI than height (0.100 (95% CI, 0.057–0.142) versus 0.028 (0.007–0.048) in UKB1; 0.245 (0.130–0.360) versus 0.062 (0.009–0.115) in BBJ1), replicating a previous report in the European population38 and suggesting a shared G×E architecture across populations for the anthropometric traits. Other traits with significant G×E heritability showed ratios ranging from 0.03 to 0.52, indicating heterogeneity in G×E contributions across traits (Supplementary Table 21). Notably, G×E heritability across quantitative traits was significantly correlated between biobanks (Spearman’s ρ = 0.41, P = 0.011; Fig. 3d), suggesting moderately concordant G×E contributions across populations.

Fig. 3: Genome-wide consistency of G×E interactions across populations.

a, G×E heritability (h2) in UKB1, aggregated across all environmental factors (Methods). Asterisks denote FDR < 0.05. BUN, blood urea nitrogen; DL, dyslipidaemia; MAP, mean arterial pressure; WBC, white blood cells. b, As in a, for BBJ1. c, Heritability of G×E interactions and marginal effects for traits with significantly positive G×E heritability in UKB1 (left) and BBJ1 (right). d, Scatter plot of G×E heritability in UKB1 and BBJ1. Regression line estimated by Deming regression to account for estimation errors in both x and y axes. P values were estimated with a two-sided Spearman’s rank correlation test. For a–d, data are presented as estimated values with 95% confidende interval (see Supplementary Table 1 for the sample sizes). e, Significant genetic correlations (Rg, FDR < 0.05) of marginal effects (grey lines) and G×E interactions (coloured lines), estimated in UKB1. The widths of lines represent the strength of genetic correlations; dashed lines represent negative genetic correlations. f, As in e, for BBJ1.

We next estimated the cross-trait correlation of G×E interactions. Significant correlations were observed for 20 and 29 trait–environment pairs in UKB1 and BBJ1, respectively (FDR < 0.05; Supplementary Table 22). Although marginal genetic correlations formed a single cluster, G×E correlations were clustered by trait categories (Fig. 3e,f). Notably, the same environmental factors mediated G×E correlations across populations: current smoking in liver-related traits; sex and dietary consumption in blood pressure traits; and age and sex in renal-related traits. These trait–environment relationships were consistent with known epidemiology and recovered without previous clinical input, suggesting that the trait–environment relationships were embedded in the genome-wide G×E architecture.

Unfiltered approach for G×E detection

Past studies have often limited G×E analyses to prefiltered variants to reduce multiple testing burden15,18. A common approach is variance quantitative trait loci (vQTL) analysis, which tests associations between genotypes and phenotypic variance without requiring environmental measurements13. In Supplementary Note 9, we examined overlaps between G×E interactions and vQTL. G×E loci were 14.6-fold enriched for vQTL compared with GWAS loci, supporting vQTL as an effective prefiltering strategy13. However, vQTL missed most G×E interactions (54.8% in UKB1 and 80.6% in BBJ1; Supplementary Table 23) and their detection was sensitive to phenotype normalization. Moreover, vQTL heritability did not correlate with G×E heritability across traits. These results underscore the necessity of using environmental data explicitly for comprehensive G×E detection.

Influence on polygenic prediction

Polygenic score (PGS)-based disease risk prediction is actively explored, but environmental differences within and across populations can reduce its prediction accuracy for specific traits11,39, potentially exacerbating health disparities. This might affect other traits in general, considering the G×E heritability for broad trait categories. To systematically assess environmental effects on polygenic prediction, we stratified the discovery and replication cohorts by environments into two groups (for example, ever-smokers and never-smokers; younger and older halves of the group). We performed GWAS and constructed PGS within individual strata of the discovery cohorts, and evaluated their prediction accuracy in the strata of the replication cohorts (Fig. 4a).

Fig. 4: Genome-wide properties of G×E interactions.

a, Overview of PGS construction and evaluation of within-population prediction accuracy and cross-population portability. b,c, Examples of within-population prediction accuracy (R2): PP in UKB stratified by age (b) and uric acid in BBJ stratified by drinking status (c). Asterisks represent FDR < 0.05. d, BMI distribution stratified by G×E-PGS, and that from marginal PGS. Boxplots show median, quartiles and 1.5× interquartile range whiskers. N = 30,683 and 40,226 for male and female participants, respectively. e, UMAP of the Tabula Sapiens scRNA-seq data. We used the 10x Genomics subset and subsampled 100 cells per tissue–anatomy–cell type combination (Methods). Cell types defined in the original study. f,g, Single-cell associations with PP genetics in the younger (f) and older (g) groups. h, UMAP of the monkey artery scRNA-seq data. Cell types defined in the original study. i, Difference in single-cell associations between the older and younger groups.

Among the 26 trait–environment–biobank triplets with significantly positive G×E heritability, 20 exhibited significant intra-population differences in prediction accuracy in at least one stratum (FDR < 0.05; Fig. 4b,c, Extended Data Fig. 9a and Supplementary Table 24). Prediction accuracy was generally the highest when applied to the same environmental group at PGS construction. Excluding one related to a UKB1-specific environment cluster (BMI–fish-and-vegetable), 11 out of 25 triplets also showed significant differences in cross-population portability (Extended Data Fig. 9b–d), although the prediction accuracy was generally attenuated.

Polygenic scores constructed from G×E interactions (G×E-PGS)40 consistently significantly explained phenotypic variance in independent cohorts (11 out of 25 (9 out of 25) trait–environment pairs within (across) populations; Supplementary Table 25). Notably, a G×Sex-based PGS constructed in BBJ1 successfully stratified BMI in opposite directions between sexes in J-MICC/HERPACC, capturing the polygenic architecture of sex differences (Fig. 4d). This stratification was apparent even for each sex and among individuals with similar marginal PGS, suggesting that PGS extended to two dimensions could enhance phenotype prediction. Indeed, a model incorporating G×E-PGS improved BMI prediction accuracy by 16% over a model without G×E-PGS (R**2 = 0.128 versus 0.110). By contrast, gains in prediction accuracy were modest for other trait–environment pairs (Supplementary Table 25). Larger sample sizes and methods tailored for G×E-PGS construction are warranted to fully realize the potential of G×E interactions in precision medicine.

Collectively, these observations demonstrated that environmental factors systematically impacted intra- and cross-population PGS prediction accuracy, and incorporating G×E interactions could enhance genetic risk prediction of human complex traits.

Aging shift of pulse pressure genetics

As genome-wide G×E architecture recapitulated epidemiologically plausible trait–environment relationships, we reasoned that G×E interactions could uncover biological mechanisms underlying the genetic dynamics of complex traits. We focused on G×Age interactions for pulse pressure, given their strong signals: all four G×E loci in UKB1 were driven by G×Age, G×Age heritability was significantly positive and PGS prediction accuracy varied across age groups (Supplementary Tables 6 and 21, and Fig. 4b).

We divided BBJ1 and UKB1 into two equal-sized age groups and conducted GWAS for pulse pressure within each group. Cross-population meta-analyses of MAGMA gene-set analyses41 revealed that vascular smooth muscle contraction was enriched in younger individuals and cellular senescence enriched in older individuals (Extended Data Fig. 9e). When projecting polygenic effects onto tissue-wide scRNA-seq data from Tabula Sapiens7,42, blood vessel cell types were significantly associated with pulse pressure in both age groups, whereas their relative strength of associations differed by age (Fig. 4e–g and Supplementary Table 26). To examine this closely, we repeated the analysis using a scRNA-seq dataset of monkey arteries43 (Fig. 4h). In younger individuals, genetic effects were associated with smooth muscle cells (P = 8.3 × 10−3 and 9.4 × 10−3 for the two subtypes), whereas in older individuals they were associated with a subgroup of coronary endothelial cells (P = 2.3 × 10−3) (Fig. 4i and Extended Data Fig. 9f,g). As endothelial cells are central to vascular senescence and atherosclerosis[44](https://www.nature.com/articles/s41586-025-10054-6#ref-CR44 “Jia, G., Aroor, A. R., Jia, C. & Sowers, J. R. Endothelial cell senescence in aging-related vascu