Introduction

Genome-wide association studies (GWAS) have revolutionized our understanding of the genetic basis of various human conditions by identifying markers associated with diseases and traits1,2,3. The GWAS catalog (https://www.ebi.ac.uk/gwas/) is a public database summarizing information from genetic associations for numerous traits, including chronic, common adult diseases, such as type 2 diabetes, cardiovascular disease, or psychopathology4. It provides evidence for the influence of hundreds to thousands of single-nucleotide polymorphisms (SNPs) on these diseases, each showing small individual effects. Exposure to adverse environmental conditions early in life (e.g., trauma, emotional abuse, neglect, and poverty) also influences adult health, increasing the risk for a wide range of outcomes, such as major depression and other psychopathologies, ischemic heart conditions, type 2 diabetes, and comorbidities5,6,7,8. While these associations apply at the population level, individuals differ in their susceptibility, and not all those exposed to adversity develop chronic adult diseases9. Indeed, measures of exposure to adversity alone cannot distinguish between individuals who will or will not experience later physical or mental health problems9. There is now considerable evidence for the importance of heritable genetic variation as a major source of susceptibility to environmental exposure10.

Recent methodologies, such as genome-wide environment interaction studies (GWEIS), examine the interaction between genetic and environmental factors, thereby enhancing the identification of SNP effects that depend on environmental exposure. GWEIS identified interactive SNPs with lifestyle factors or trauma exposure to predict cardiovascular traits, such as blood pressure and lipids11 or depression and neuroticism12,13. These studies also indicate only a modest correlation between main genetic effects and interaction effects11, reflecting the complex interplay between genetic and environmental influences in defining variability in health, well-being, and productivity over the life span. These findings also raise an important yet unresolved question: how do the genetic variants associated with disease risk compare with those that moderate the impact of environmental adversity?

To directly address this gap, we conducted GWASs in 180,129 individuals from the UK Biobank to identify genetic variants associated with the risk of common adult chronic physical and mental disorders. We then performed GWEISs in the same population and under the same analytical framework to characterize genetic variants associated with disease risk in response to early adversity (22,372 individuals reported feeling hated or unloved during childhood). We hypothesize that a distinct set of genetic variants is associated with disease risk or the moderation of early adversity on the same outcomes. We show that variants associated with disease risk do not overlap with those identified in gene–environment interaction analyses across physical and mental health disorders, and that SNPs identified in GWEIS fall outside the statistically significant regions of the corresponding GWAS Manhattan plots. We further show that heritability estimates are detectable for the main genetic analyses but not for the interaction analyses, and are higher for physical diseases than for mental health conditions. Together, these findings indicate that distinct genetic architectures underlie disease susceptibility and genetic moderation to early life adversity.

Materials and methods

Ethics approval

Informed written consent was obtained from all participants of the UK Biobank cohort. Approval for the UK Biobank was obtained by the North West Multicentre Research 580 Ethics Committee (REC reference 11/NW/0382; www.ukbiobank.ac.uk/ethics/), the National Information Governance Board for Health and Social Care, and the Community Health Index Advisory Group. This research was conducted using the UK Biobank Resource under Application Number 41975. Local ethics approval was obtained from the Research Ethics Board – Mental Health and Neuroscience subcommittee based at the Centre intégré universitaire de santé et de services sociaux de l’Ouest-de-l’Île-de-Montréal Research Ethics Board under the number 2022-371, IUSMD 21-73. Consenting participants provided baseline information, answered questions, and had measurements and biological samples collected. All ethical regulations relevant to human research participants were followed in accordance with the Declaration of Helsinki.

Study population

The UK Biobank (UKB) cohort is a population-based cohort of 502,543 individuals aged 37–73 years, recruited from 23 assessment centers across the United Kingdom between 2006 and 2010. Participants provided both phenotypic and genotypic data. Genotype data were available for 487,409 subjects. We excluded participants who had withdrawn consent, exhibited discrepancies between genetic and self-reported sex, or were identified as outliers for heterozygosity adjusted by genetics PCs (>0.1903) and missing rates (>0.05), which identify the genotyping samples of poor quality (ID22027). Also, we retained only those subjects who identified themselves as “Caucasians” (ID22006) and had data available on Adversity (ID20487, ID20489). Additionally, we excluded variants with a minor allele frequency (MAF) < 0.01 (i.e., <1%), imputation accuracy INFO score <0.6, as well as duplicated or ambiguous SNPs. After applying these initial quality-control filters based on UKB-provided variant information, the dataset included 180,129 individuals and 7,288,026 genetic variants. Information about the genotyping protocol, quality control (QC), and imputation procedures is available elsewhere14, and was applied in our previous studies using the UKB7,15. Because REGENIE16 recalculates allele frequencies within the analytic sample, MAF was subsequently re-estimated in the final study sample (n = 180,129), and variants with MAF < 0.01 were additionally excluded, resulting in a final set of 7,163,199 genetic variants. The population structure was evaluated in the UKB using the fastPCA algorithm17 for principal components (PC) analysis, and the first forty PCs were included as covariates in all analyses to account for population stratification.

Reported early adversity

Two assessments, the Mental Health and Mental Wellbeing Online Follow-up, included questions from the childhood trauma screener (CTS-5)18,19,20, a brief version of the childhood trauma questionnaire (CTQ), which assesses five subtypes of childhood maltreatment: emotional, physical, and sexual abuse (acts of commission), and emotional and physical neglect (acts of omission). Participants responded to one item per maltreatment subtype using a 5-point Likert scale ranging from 1 (never true) to 5 (very often true). Childhood emotional abuse and neglect reflect persistent experiences of emotional rejection, insufficient emotional support, and a lack of perceived care during childhood21. Consistent with this conceptualization and prior validation work on the CTQ and CTS, items capturing feelings of being hated or unloved are central indicators of emotional abuse and neglect22,23,24. In the present study, we considered that participants were exposed to adversity if they reported feeling hated during childhood by a family member sometimes, often, or very often (“when I was growing up… I felt that someone in my family hated me,” ID20487, ID29078) or if they reported never or rarely feeling loved during childhood (“when I was growing up… I felt loved,” ID20489, ID29076) (Table 1)7. Early adversity exposure was then modeled as a binary variable (exposed vs. non-exposed). We used the first assessment (ID20487, ID20489) when available and the second assessment (ID29078, ID29076) otherwise. Using this classification, 22,372 individuals reported exposure to emotional abuse and neglect, and 157,757 reported no such exposure (Supplementary Table 1). We focused on childhood emotional abuse and neglect based on evidence from our recent UK Biobank study showing that these forms of maltreatment are robustly associated with adverse physical and mental health outcomes7, and because of the increasing recognition that these types of abuse are particularly harmful, in some cases showing equal or greater long-term impact than other forms of childhood maltreatment25.

Table 1 The childhood trauma screener (CTS) items and the criteria used to classify exposure to childhood adversity
Full size table

Phenotypes (common chronic diseases)

For each phenotype, cases were defined as individuals with at least one relevant diagnosis recorded during hospital inpatient admissions using the International Classification of Diseases, 10th Revision (ICD-10; ID41270): cardiovascular disorders (ischaemic heart disease [I20–I25], atherosclerosis [I70], and cerebral atherosclerosis [I67.2]); non-insulin-dependent diabetes mellitus (type 2 diabetes, E11); mood disorders (manic episode [F30], bipolar affective disorder [F31], depressive episode [F32], recurrent depressive disorder [F33], persistent mood disorders [F34], other mood disorders [F38], and unspecified mood disorder [F39]); and neurotic disorders (phobic anxiety disorders [F40], other anxiety disorders [F41], obsessive-compulsive disorder [F42], reaction to severe stress and adjustment disorders [F43], dissociative disorders [F44], somatoform disorders [F45], and other neurotic disorders [F48]). Controls were defined separately for each outcome as individuals without a diagnosis of the corresponding disease. Participants diagnosed with more than one condition were considered cases in all relevant GWAS/GWEIS analyses. Accordingly, individuals could serve as controls for one outcome while being classified as cases for another. The distribution of diagnoses in the sample was as follows: no diagnoses, 159,531 (88.56%); one diagnosis, 17,329 (9.62%); two diagnoses, 2910 (1.62%); three diagnoses, 322 (0.18%); and four diagnoses, 37 (0.02%). A detailed descriptive table of the sample sizes for the four outcomes is depicted in the Supplementary Table 1.

GWAS and GWEIS statistical models

We used REGENIE16, a program for whole-genome regression modeling in large GWAS that accounts for relatedness between individuals, to test the main effect of SNPs (GWAS) and the interaction between SNPs and early-life adversity (GWEIS) on the phenotypes. All models were adjusted for genetic sex (ID22001), age at recruitment (ID21022), genotyping array (ID22000), assessment center (ID54), and the first forty genetic principal components (ID22009) to account for population stratification, as residual population structure can still be present even within ancestrally homogeneous groups, and the inclusion of principal components is standard practice in UK Biobank–based genetic analyses26. The following models were performed:

GWAS: phenotype ~ SNP + sex + age + genotyping_array + assessment_center + PC1 +… + PC40

GWEIS: phenotype ~ SNP + adversity + SNP × adversity + sex + age + genotyping_array + assessment_center + PC1 +… + PC40

We conducted a sensitivity analysis to assess the robustness of SNP × adversity interaction effects in the GWEIS to the inclusion of all other covariates × adversity interaction terms. To this end, we re-estimated the SNP × adversity interaction effects for the top 100 interaction effects with the genotyped SNPs (INFO = 1) using generalized estimating equations (GEE), which takes into consideration relatedness between the subjects. The information about family relatedness was not collected in the UK Biobank; hence, we used the kinship coefficient (ID22012) to infer the family structure, attributing to one family first-degree relatives (parent-offspring, twins, and siblings) with a kinship coefficient greater than 2−5/2 27. All subjects with a kinship coefficient less than the set cut-off of 2−5/2 were assigned to single-member families (i.e., a family consisting of a single subject). For the remaining subjects with the first-degree relatedness, we applied the following procedure: for each related pair, both subjects and all other subjects with the first-degree relationship to either of the two subjects were assigned to one family.

Comparison between GWAS and GWEIS findings

For the comparison analysis, we retained independent SNPs based on the linkage disequilibrium (LD) structure of the European reference panel from the 1000 Genomes Project28. To investigate the overlap between SNPs and genes in GWAS and GWEIS for each of the four outcomes, we selected the SNPs with a nominal significance (p < 5 × 10−5) in either GWAS or GWEIS for comparison.

We then selected the top 20 GWEIS SNPs from each condition and compared their p-values with those identified in our GWAS of the same population, and in published GWAS from the literature using GWAS catalog29, a publicly available resource containing more than 7500 publications and 155485 full summary statistics.

Linkage disequilibrium (LD) score regression (LDSC) was applied to estimate SNP-based heritability (h2SNP) and genetic correlation (rg) between GWAS and GWEIS findings for each of the four phenotypes30,31. To quantify polygenic overlap irrespective of genetic correlation and using GWAS summary statistics, cross-trait MiXeR was applied32.

Statistics and reproducibility

We conducted GWAS and GWEIS using REGENIE (v4.1)16, which accounts for population structure and relatedness in large-scale genetic analyses. Several steps were taken to minimize potential sources of bias. Analyses were restricted to individuals of Caucasian ancestry and adjusted for relevant demographic, technical, and genetic covariates, including principal components to account for population stratification. Standard genotype QC procedures were applied, and relatedness between participants was accounted for using whole-genome regression models.

Genome-wide significance was defined using the conventional threshold of p < 5 × 10−8. For comparative analyses between GWAS and GWEIS results, a nominal threshold of p < 5 × 10−5 was used to assess overlap between signals. A post-hoc sensitivity analysis based on the observed range of standard errors from the completed GWAS/GWEIS models was conducted. Specifically, we estimated the minimum detectable odds ratio across different standard errors, statistical significance thresholds (α = 0.05, 5 × 10−5, and 5 × 10−8), and statistical power levels (60, 70, and 80%) using an approximation based on logistic regression effect size estimation (Supplementary Fig. 1).

SNP-based heritability and genetic correlation were estimated using linkage disequilibrium score regression (LDSC)30,31, and polygenic overlap was assessed using MiXeR32. All statistical analyses and visualizations were conducted in R (v4.1.1)33. The UK Biobank data used in this study are available through application to the UK Biobank resource, and genome-wide summary statistics generated in this study are publicly available (see “Data availability”). Detailed descriptions of the statistical models are provided above.

Results

GWAS and GWEIS significant SNPs

GWAS detects statistically significant associations based on stringent multiple-testing correction, which can identify significant SNPs despite environmental heterogeneity. However, SNPs whose effects depend on environmental exposure (i.e., G × E interactions) are generally less likely to reach genome-wide significance in GWAS, as their effects may be attenuated when averaged across individuals exposed to heterogeneous environments. We hypothesized that SNPs identified in GWEIS would fall outside the statistically significant regions of the corresponding GWAS Manhattan plots (Fig. 1). To test this, we generated GWAS and GWEIS Manhattan plots for the four conditions, with SNPs displayed in gray for GWAS and in green for GWEIS. The red horizontal lines indicate the significant threshold of p < 5 × 10−8 and the blue lines the suggestive, nominal threshold of p < 5 × 10−5 (Fig. 2). Using the p < 5 × 10−8 threshold, fifty-four SNPs were identified in the GWAS of cardiovascular diseases, and seventy-four for type 2 diabetes (Fig. 2a, b); no significant SNPs were observed for the GWAS of mood disorders or neurotic disorders (Fig. 2c, d), and none in the GWEIS analyses.

Fig. 1: Conceptual illustration of genome-wide association studies (GWAS).
Fig. 1: Conceptual illustration of genome-wide association studies (GWAS).
Full size image

GWAS are conducted in samples composed of cases and controls (individuals with or without the outcome of interest). Both groups include individuals who may have experienced environmental adversity of varying types and intensity, illustrated by the yellow symbols in the left panel. In GWAS, the association between each SNP and the outcome is tested across all individuals regardless of environmental exposure. The Manhattan plot displays SNPs ordered along the genome according to their chromosomal position (x-axis), with the strength of association shown as −log10(p-values) on the y-axis. SNPs that remain significant after correction for multiple testing appear above the genome-wide significance threshold and represent variants with strong main genetic effects on the phenotype. Variants whose effects depend on environmental exposure (SNP × environment interactions) may not show strong main effects in GWAS because environmental heterogeneity reduces their average effect across individuals. Such variants may therefore fall below the genome-wide significance threshold in GWAS but can be detected in genome-wide environment interaction studies (GWEIS). Created with BioRender.com.

Fig. 2: GWAS and GWEIS Manhattan plots for physical and mental health outcomes.
Fig. 2: GWAS and GWEIS Manhattan plots for physical and mental health outcomes.
Full size image

Manhattan plots from genome-wide association studies (GWAS; gray) and genome-wide environment interaction studies (GWEIS; green) are shown for a cardiovascular disease, b type 2 diabetes, c mood disorders, and d neurotic disorders. Genome-wide analyses were conducted using REGENIE, which implements a whole-genome regression framework. Association tests were two-sided, testing either SNP main effects (GWAS) or SNP × early-life adversity interaction effects (GWEIS). Correction for multiple comparisons was addressed using the conventional genome-wide significance threshold (p < 5 × 10−8), indicated by the red horizontal line. The blue horizontal line represents a suggestive (nominal) threshold (p < 5 × 10−5). SNPs reaching the nominal significance threshold (p < 5 × 10−5) in the GWEIS analyses are highlighted in green in the corresponding GWAS Manhattan plots. Consistent with the hypothesis illustrated in Fig. 1, these SNPs do not reach the upper significance range of the GWAS Manhattan plot, indicating limited overlap between main genetic effects and gene–environment interaction effects. All analyses were performed in the same sample of 180,129 UK Biobank participants. Cardiovascular disorders: 10,655 cases and 169,474 controls, type 2 diabetes: 5648 cases and 174,481 controls, mood disorders: 4761 cases and 175,368 controls, neurotic disorders: 3199 cases and 176,930 controls.

We next selected all SNPs at the nominal level of significance in the GWEIS (at p < 5 × 10−5), and overlaid them in green on the corresponding GWAS plots (Fig. 2). Consistent with our hypothesis, the nominally significant GWEIS SNPs were located in non-significant regions of the GWAS plots, consistent across all four conditions (Fig. 2).

The genomic inflation factor (λGC) was calculated for each GWAS phenotype using the full set of tested SNPs. λGC values ranged from 1.034 to 1.101, with the highest value observed for type 2 diabetes, consistent with a higher polygenic signal in this GWAS. Visual inspection of the QQ plots did not indicate systematic bias, with deviations from the null expectation confined to the tail of the distribution, supporting a predominantly polygenic rather than confounded origin of the signal (Supplementary Fig. 2).

In GWEIS sensitivity analyses, we re-estimated SNP × adversity interaction effects for the top 100 GWEIS SNPs after including all covariate × adversity interaction terms in the model. The resulting SNP × adversity interaction estimates were highly consistent with those obtained in the original GWEIS analyses. Across all four outcomes, interaction beta estimates from the GEE models were strongly correlated with those from the REGENIE analyses (Pearson’s r = 0.99), indicating that inclusion of covariate × adversity interaction terms had minimal impact on the estimated SNP × adversity interaction effects.

Comparison between GWAS and GWEIS findings

For all analyses in this subsection, SNP lists were restricted to variants included in the LD Hub reference panel (~1 million SNPs) rather than the full GWAS set (7,163,199 SNPs). We analyzed the overlap between the nominally significant SNPs from the GWAS and GWEIS for the different health outcomes at the p < 5 × 10−5 threshold. No overlap was observed at the SNP or gene level between GWAS and GWEIS for any of the outcomes (Supplementary Data S1 and Fig. 3).

Fig. 3: Overlap between SNPs and genes identified in GWAS and GWEIS analyses across the four health outcomes.
Fig. 3: Overlap between SNPs and genes identified in GWAS and GWEIS analyses across the four health outcomes.
Full size image

Overlap analyses were conducted between SNPs and genes reaching a nominal significance threshold (p < 5 × 10−5) in the GWAS for four chronic diseases and those identified in the GWEIS (SNP-by-early adversity model) for the same outcomes. Cardiovascular disorders: 10,655 cases and 169,474 controls, type 2 diabetes: 5648 cases and 174,481 controls, mood disorders: 4761 cases and 175,368 controls, neurotic disorders: 3199 cases and 176,930 controls.

We then selected the top GWEIS SNPs from each condition and compared their p-values with those identified in our GWAS and in previously published GWAS available in the GWAS catalog29. The top SNPs in the GWEIS model exhibited non-statistically significant p-values in our own GWAS for the same population and condition (Tables 2 and 3). Additionally, only 8 of the top GWEIS SNPs showed associations in previously published GWAS at p < 5 × 10−5, and none were associated with the same trait. From these, only 3 were previously identified in GWAS at p < 5 × 10−8: from cardiovascular disease GWEIS, rs1003645 was previously statistically significant on a GWAS for Ck-beta-8-1 levels (related to inflammatory responses)34, neonatal cytokine levels35, and height36. From the same cardiovascular disease GWEIS, rs1617208 was previously identified in a GWAS for blood protein levels34 (Table 2). For type 2 diabetes and mood disorders GWEIS, no SNPs showed significant associations in published GWAS at p < 5 × 10−8. In contrast, one neurotic disorders-related GWEIS SNP (rs10936142) was identified in a prior GWAS for diastolic blood pressure37 (Supplementary Data S2 and Table 3).

Table 2 Comparison of the top GWEIS SNPs associated with physical health disorders and corresponding GWAS findings from the same cohort and the GWAS catalog
Full size table
Table 3 Comparison of the top GWEIS SNPs associated with mental health disorders and corresponding GWAS findings from the same cohort and the GWAS catalog
Full size table

LD score regression (LDSC) revealed no detectable SNP-based heritability for the GWEIS. For the GWAS, SNP-based heritability estimates were greater for physical diseases (2.9% for cardiovascular disease [h2 = 0.0289, SE = 0.0035] and 2.3% for type 2 diabetes [h2 = 0.0237, SE = 0.003]) than for mental conditions (0.7% for neurotic disorders [h2 = 0.0071, SE = 0.0029] and 0.9% for mood disorders [h2 = 0.0089, SE = 0.0027]). Genetic correlation between GWAS and GWEIS findings could not be estimated because SNP-based heritability for the GWEIS was near zero.

To estimate polygenic overlap of GWAS and GWEIS, we applied MiXeR. For the GWEIS, the MiXeR model yielded negative AIC values, indicating insufficient statistical support to justify the more complex MiXeR model relative to LDSC. This suggests limited power in the GWEIS data to reliably estimate the polygenic architecture parameters required by MiXeR. Consistent with LDSC results, MiXeR could not reliably estimate genetic correlation between GWAS and GWEIS due to insufficient statistical support for the GWEIS.

In sum, the genetic background associated with disease risk identified by GWAS is distinct from the genetic background that modifies the effects of early adversity, as captured by GWEIS.

Discussion

In this study, we observed that genetic variants linked to main genetic associations with chronic adult diseases, as identified in the GWAS, differ substantially from those that modify the association between environmental exposures and adult disease (identified in the GWEIS). These two types of analyses identify distinct components of disease risk. It is important to highlight that we used the same dataset for both the GWAS and GWEIS studies. By analyzing the same cohort for both genetic associations and gene–environment interactions, we minimized variability that might arise from population differences. Overall, only a limited subset of top GWEIS variants overlapped with genome-wide significant loci reported in the GWAS catalog. In the rare instances in which the overlap was observed, it occurred for distinct traits rather than for the same disease outcomes, even when biologically related traits could be identified (e.g., a SNP identified in the cardiovascular disease GWEIS is also significant in a GWAS for Ck-beta-8-1 levels, a cytokine predominantly expressed in the lung and liver).

We demonstrated using Manhattan plots that SNPs identified in the GWEIS were located in non-significant regions of the corresponding GWAS plots. A Manhattan plot displays chromosomal position on the x-axis and −log10(p-values) on the y-axis, with genome-wide significant loci appearing above the multiple testing threshold. GWAS detects SNPs through their main effects across individuals, without explicitly modeling gene–environment interactions. As a result, genetic effects are estimated across the entire sample, regardless of environmental exposure. Because environmental exposures are distributed across both cases and controls and across all genotypes at each SNP, their effects are effectively averaged within each genotype group (e.g., AA, AB, and BB). When the effect of a SNP depends on environmental context, this averaging can attenuate or obscure associations in GWAS, particularly when effects differ in magnitude or direction across exposure groups. In contrast, GWEIS explicitly models these interactions, allowing detection of context-dependent genetic effects that may not be apparent in GWAS. This framework explains why GWEIS-identified SNPs fall outside the significant regions of GWAS Manhattan plots and highlights the importance of interaction-based approaches for identifying environmentally dependent genetic effects.

Additionally, SNPs identified through statistically significant G × E interaction effects in GWEIS are often not detected in GWAS because their effects may vary in direction or magnitude depending on the environmental context. In such cases, opposing or context-dependent effects can cancel out when averaged across individuals, attenuating their association in GWAS. This pattern is consistent with the differential susceptibility framework38. Consistent with this interpretation, we observed no detectable SNP-based heritability in the GWEIS, reflecting the context-dependent nature of these effects.

Similar to our study, Battram and colleagues39 evaluated whether the genes or gene ontology terms identified by GWAS and epigenome-wide association studies (EWAS) overlap, investigating fifteen complex traits. EWAS identifies variations in DNA methylation in response to environmental stimuli. Among the 15 EWAS examined, 11 studies identified gene sets that did not overlap with those from the corresponding GWAS beyond what would be expected by chance, and only one study showed overlap for diastolic blood pressure. These findings suggest that GWAS and EWAS are also capturing distinct gene sets associated with complex traits, a result that is aligned with our current findings indicating a lack of overlap between GWAS and GWEIS results39.

Consistent with the overall lack of overlap between GWAS and GWEIS findings, the top SNPs identified in the interaction analyses showed non-significant p-values in the corresponding GWAS conducted in the same population. Moreover, only a subset of these variants had been previously reported in published GWAS, and none were directly associated with the same disease outcomes examined here. Instead, several of the identified variants were associated with intermediate phenotypes, such as inflammatory markers, circulating proteins, or blood pressure. These findings suggest that gene–environment interaction loci may act through biological processes that influence an individual’s sensitivity to environmental exposures, rather than directly contributing to baseline disease risk40,41. In this context, methods that incorporate functional genomic features or leverage biologically informed pathways responsive to environmental variation may offer a more effective framework for identifying gene–environment interactions42,43. Our findings further support the idea that genetic susceptibility to disease and genetic moderation to early life adversity may arise from partially distinct biological mechanisms13.

Our study has limitations. First, statistical power remains an important consideration. Although our analyses were conducted in a large and well-characterized cohort (N = 180,129), GWEIS are inherently less well powered than traditional GWAS because interaction effects are typically smaller and require substantially larger sample sizes to detect reliably. Consequently, some SNP × adversity interaction effects may have remained undetected, particularly those with modest effect sizes. This limited power may also contribute to the absence of SNPs reaching the conventional genome-wide significance threshold (p < 5 × 10−8) in the interaction analyses and to the limited overlap observed between GWAS and GWEIS signals. While the use of REGENIE allowed us to efficiently account for relatedness and maximize the available sample size, power constraints remain an important consideration when interpreting null findings. In addition, although analyses were restricted to individuals of Caucasian ancestry, residual population structure may persist within the UK Biobank14,44,45; to mitigate this, we included the first 40 genetic principal components, a widely adopted and conservative approach in large-scale UK Biobank analyses. Second, early-life adversity was assessed retrospectively using self-reported measures in the UK Biobank, which may introduce recall bias and misclassification of exposure. Prospective and retrospective assessments of childhood maltreatment are known to show limited concordance, suggesting that some cases of early adversity may be under- or over-reported depending on the assessment approach46. Prospective assessments may capture only the most severe cases, whereas retrospective reports may identify a broader range of experiences. Nevertheless, although these measures are not interchangeable, both prospective and retrospective assessments of childhood adversity have been consistently associated with adult physical and mental health outcomes9.

Overall, our findings help explain the limited success of studies using GWAS-derived polygenic scores to investigate gene–environment interactions in mental and physical disorders47,48. The lack of overlap between GWAS and GWEIS results indicates that genetic variants identified through main effects do not fully capture context-dependent genetic influences. This has important implications, as many downstream applications (including risk prediction, biological interpretation, and therapeutic target identification) are currently based primarily on GWAS findings. While GWEIS provides a framework to identify variants that moderate the effects of environmental exposures, our results suggest that relying solely on GWAS-derived signals provides an incomplete view of genetic risk. Our findings underscore the need for approaches that explicitly account for environmental context to better characterize the context-dependent genetic architecture of complex diseases.