Abstract
Cardiovascular disease (CVD) is influenced by genetic and environmental factors. Childhood maltreatment is associated with CVD and may modify genetic susceptibility to cardiovascular risk factors. We used genetic and phenotypic data from 100,833 White British UK Biobank participants (57% female; mean age = 55.9 years). We regressed nine cardiovascular risk factors/diseases (alcohol consumption, body mass index [BMI], low-density lipoprotein cholesterol, lifetime smoking behaviour, systolic blood pressure, atrial fibrillation, coronary heart disease, type 2 diabetes, and stroke) on their respective polygenic scores (PGS) and self-reported exposure to childhood maltreatment. Effect modification was tested on the additive and multiplicative scales by including a product term (PGS*maltreatment) in regression models.
On the additive scale, childhood maltreatment accentuated the effect of genetic susceptibility to higher BMI. Individuals not exposed to childhood maltreatment had an increase in BMI of 0.12 SD (95% CI: 0.11, 0.13) per SD increase in BMI PGS, compared to 0.17 SD (95% CI: 0.14, 0.19) in those exposed to all types of childhood maltreatment. On the multiplicative scale, similar results were obtained for BMI though these did not withstand to Bonferroni correction. There was little evidence of effect modification by childhood maltreatment in relation to other outcomes, or of sex-specific effect modification.
Our study suggests the effects of genetic susceptibility to a higher BMI may be moderately accentuated in individuals exposed to childhood maltreatment. However, gene*environment interactions are likely not a major contributor to the excess CVD burden experienced by childhood maltreatment victims.
Introduction
Cardiovascular disease (CVD) is the leading cause of death worldwide.1 Many CVD subtypes have complex aetiologies, affected by both genetic and environmental factors.2,3 This suggests that genetic and environmental factors may interact with each other and modify susceptibility to disease. Childhood maltreatment increases the risk of a number of CVD subtypes and related diseases, including coronary heart disease (CHD), stroke and type 2 diabetes mellitus (T2DM).4,5 Evidence also suggests there is a dose-response relationship between childhood maltreatment and CVD, with disease risk increasing according to the number of types of maltreatment experienced by an individual.6–13 While this growing body of evidence indicates childhood maltreatment is associated with higher CVD risk, it is unknown whether this risk is disproportionate in genetically susceptible individuals.
Single nucleotide polymorphisms (SNPs) detected in genome-wide association studies (GWAS) have been used to construct polygenic scores (PGSs) for a trait or disease. These estimate the genetic risk carried by an individual by adding the effects of all relevant variants in the genome. The polygenic nature of many diseases, including CVD, suggests these scores may be more informative when investigating interactions or effect modification by environmental factors, in comparison to using a single associated gene.14 PGSs have previously been employed to assess gene*environment interactions or effect modifications for cardiovascular traits, including for educational achievement, neighbourhood and obesogenic environments.15–17 Identifying these effects may help to understand the relationship between early life environmental exposures and CVD risk in adult life.
This study aims to investigate whether childhood maltreatment modifies genetic susceptibility to a range of cardiovascular risk factors and CVD subtypes, using data from the UK Biobank cohort. Where two risk factors are associated with the outcome, evidence of effect modification is expected on either the additive or the multiplicative scale, or on both scales.18,19 Therefore, in this analysis we investigate both scales. Of greatest relevance to public health is the direction and magnitude of any interaction. We hypothesise that childhood maltreatment accentuates the effect of polygenic predisposition to CVD subtypes and risk factors. We also investigate whether any potential effect modification differs according to sex, as some studies have detected differences in the association between childhood maltreatment and cardiovascular risk13,20–23
Methods
UK Biobank
UK Biobank recruited 503,317 adults aged 37-73 years at baseline, between 2006 and 2010.24 At baseline, data was collected on health and lifestyle information, physical measurements and biological samples. Some follow-up questionnaires and clinics have taken place, including an online mental health questionnaire in 2016, which asked about childhood maltreatment. This was sent to participants who provided an email address (N=339,092) and 157,366 (46.4%) responses were obtained. This analysis includes 100,833 participants of White British ancestry who completed questions on childhood maltreatment and had genetic information. Some individuals were missing CVD risk factor information and thus sample sizes for analyses ranged from n=93,001 to n=100,833 (Table 1).
Childhood maltreatment
The types of maltreatment assessed comprised of emotional neglect, physical neglect, emotional abuse, physical abuse and sexual abuse.25,26 A summary score of childhood maltreatment was calculated as per the Childhood Trauma Screener,25 reflecting the cumulative number of maltreatment types experienced by an individual, spanning from 0 to 5 (Table S1).
Cardiovascular risk factors and diseases
This study included 3 CVD subtypes and 6 risk factors for CVD. Risk factors were considered if there is evidence for a causal association with CVD from randomised control trials (RCT) or Mendelian randomisation studies (see Table S2) - a statistical approach using randomly allocated genetic variants as instrumental variables 27 - and have GWAS information with summary statistics available.
Quantitative traits (body mass index (BMI), alcohol consumption measured as drinks per week, low-density lipoprotein cholesterol (LDL-C), lifetime smoking behaviour and systolic blood pressure (SBP) were determined at baseline through physical or biological measurements, or self-reported for alcohol and smoking behaviours.
Incidence of CVD subtypes (atrial fibrillation (AF), CHD and stroke) and of risk factor T2DM were determined through linked mortality data, hospital episode statistics and Scottish morbidity and mortality records (SMR), defined using ICD-9 and ICD-10 codes (Table S3).
The period of follow-up occurred between the date of baseline assessment (between 2006 and 2010) and the most recent available linked hospital inpatient data, from May 2017.
Polygenic scores
Summary statistics were obtained from the most recent GWAS (Table 2). Where the most recent GWAS included UK Biobank participants, the next most suitable study was used to avoid estimate inflation through sample overlap. PGSs were calculated by multiplying the number of effect alleles by the relative effect estimate of the SNP on the phenotype and summing together the alleles.
PGSs were generated at P value thresholds of <5×10−8 (genome-wide significance), <0.05 and <0.5. The greater the P value threshold for SNPs, the greater the variance explained by the score (Table S4), but the greater the chance of capturing pleiotropic effects. Resulting PGSs were standardised, reflecting a one standard deviation increase in PGS.
Full details of the derivation of PGSs, GWAS studies used and generation of phenotypic measures can be found in the Supplement.
Exclusion criteria
Quality control of UK Biobank genetic data was performed in accordance with the MRC-IEU Quality Control pipeline, as previously described.28 In brief, participants with sex mismatch (comparing genetic sex with reported sex) and those with sex chromosome aneuploidy were excluded from the study. Participants of White British ancestry were those who self-reported as “White British” and supported by genetic principal component (PC) analysis. Related participants were excluded according to kinship coefficients provided by the UK Biobank until no related pairs remained.28
Individuals missing age, sex, genetic data or who did not have complete information for childhood maltreatment were excluded from the sample (Figure 1). Individuals missing risk factor or disease data were excluded from relevant analyses. To prevent the detection of effects caused by reverse causality, participants were excluded if they experienced any diagnosis of AF, T2DM, CHD or stroke prior to baseline. Prevalent cases were ascertained by comparing the date of diagnosis in the linked hospital inpatient data to the date of baseline visit.
Statistical analysis
Statistical analyses were carried out using Stata16-MP. Analysis code is available at github.com/helenaurquijo/gxe_cvd_maltreatment. Study data used will be archived with UK Biobank and was carried out under project 19278.
Association of childhood maltreatment with outcomes
Multivariable linear and logistic regression models, adjusting for sex and age, were used to estimate the association between the childhood maltreatment score and each of the nine risk factors/diseases.
Effect modification of childhood maltreatment on genetic susceptibility
For continuous traits, the association of childhood maltreatment score and each PGS with the respective trait was estimated on the additive scale via multivariable linear regression. On the multiplicative scale, the association was estimated via multivariable linear regression for the natural log of the trait. For binary traits, the association of childhood maltreatment score and each PGS with the respective disease trait was estimated on the additive scale via multivariable linear regression. On the multiplicative scale, the association was estimated via multivariable logistic regression. All regressions were adjusted for sex, age and 40 genetic PCs. Continuous phenotypic measures of traits were standardised to reflect a change in SD of trait per 1 SD increase in PGS. Coefficients for binary traits reflect a change in risk per SD increase in PGS.
To test for effect modification of genetic predisposition by exposure to childhood maltreatment, the same regression models were run including a product of the PGS and maltreatment score as an interaction term. Effect modification was evaluated by the size and precision of the coefficient for effect modification, and the P value for the effect modification coefficient considering Bonferroni correction.
To address the risk of type 1 error due to multiple comparisons, a Bonferroni correction was applied by dividing the standard 0.05 P value threshold by the number of outcomes we considered in each analysis (nine), yielding a threshold of P=0.006. The use of this correction approach was conservative as the considered traits were not independent from each other.
Secondary analyses
Sex-stratified analyses were carried out, and a 3-way interaction term (i.e. PGS*maltreatment*sex) on the respective risk factor or CVD subtype was included.
Main analyses were repeated using PGSs generated with less stringent P value thresholds at 0.05 and 0.5, as described previously.
Finally, sensitivity analyses were performed by replicating the main analyses with adjustments for covariate interactions (e.g. sex*PGS, age*maltreatment…) to avoid specification error.14 Separately, maternal smoking and number of siblings (proxies of childhood socioeconomic position), which may confound associations between childhood maltreatment and CVD, were included in the adjustment.
Results
UK Biobank sample
Eligible participants had a mean age at baseline of 56 years (SD: 8.0 years) and 57% were female. Most participants (55%) reported no exposure to any type of childhood maltreatment. The most frequently reported type of maltreatment was emotional neglect (21%) and sexual abuse was the least reported (8%) (Table 1). Our study sample did not differ from the rest of the UK Biobank in terms of exposure to maltreatment, cardiovascular risk and cases of CVD.
With the genome-wide significance threshold (P<5×10−8), variance explained by the PGSs ranged between 0.06% (smoking) and 15% (systolic blood pressure) (Table S4).
Association of childhood maltreatment with cardiovascular risk factors and disease
Childhood maltreatment score was associated with higher BMI (β: 0.07 SD; 95% CI: 0.06, 0.07) and lifetime smoking, as well as with T2DM and CHD incidence (Table S5). However, childhood maltreatment was negatively associated with systolic blood pressure (β: −0.011 SD; 95% CI: −0.017, - 0.005). Childhood maltreatment was positively associated with all other outcomes, but the confidence intervals spanned the null.
Effect modification by childhood maltreatment
On the additive scale, there was little evidence that childhood maltreatment modified the effect of the PGSs for all risk factors/diseases except for BMI (Table S6 and Figure 2). The effect of genetic susceptibility to higher BMI was accentuated by childhood maltreatment score by 0.009 SD (95% CI: 0.003, 0.014; Peffect modification: 0.003). This indicates, in those unexposed to childhood maltreatment, each SD increase in the BMI PGS was associated with 0.12 SD higher BMI (95% CI: 0.11, 0.13), whilst in those with a maltreatment score of 5, each SD increase in the PGS was associated with 0.17 SD higher BMI (95% CI: 0.14, 0.19). All other interaction effect estimates were positive, with the exception of SBP, and extremely small.
On the multiplicative scale, similar yet less precise results were obtained (Table S7 and Figure 3). The effect of genetic susceptibility to higher log BMI was accentuated by childhood maltreatment, through an increase of 0.007 SD (95% CI 0.001, 0.013) per unit increase in maltreatment score. Nonetheless, the interaction estimate did not withstand Bonferroni correction (Peffect modification: 0.015). The coefficient for effect modification was in the positive direction for alcohol consumption, LDL-C, smoking, AF and stroke, and in the negative direction for SBP, CHD and T2DM. However, for all outcomes the estimates were close to the null and confidence intervals were narrow.
Secondary analyses
Sex was not found to modify the effects observed in the main analyses on either the additive or the multiplicative scale (Tables S8 and S9).
Analyses repeated with PGSs generated at P value thresholds of P≤0.05 and P≤0.5 found childhood maltreatment score to accentuate the effect of genetic susceptibility to higher BMI and greater lifetime smoking, on both the additive and multiplicative scale (Tables S10 and S11).
Analyses adjusting for covariate (age, sex and genetic PCs) interactions with childhood maltreatment score and PGSs yielded similar results, though the interaction coefficients for BMI became less precise, decreasing the strength of evidence for effect modification on the multiplicative scale (Table S12).
Analyses adjusting for maternal smoking and number of siblings yielded similar results to main analyses (Table S13).
Discussion
We found evidence that childhood maltreatment accentuates the effect of polygenic susceptibility to BMI on observed BMI but no other cardiovascular risk factors or outcomes considered. Effects were consistent across the additive and multiplicative scales, though evidence of effect modification for BMI was weaker on the multiplicative scale. Additionally, we found no evidence of sex differences in these effects.
Results in context
Few studies have investigated whether childhood maltreatment interacts with, or modifies the effect of genetic susceptibility to a range of cardiovascular risk factors/outcomes. In a candidate gene approach, Gooding and colleagues found no association between childhood maltreatment and hypertension in young adulthood, nor the presence of effect modification by the SLC64A genotype.29 In our study of adults, we found a negative association between childhood maltreatment and SBP. In a study with 161 individuals with major depression, Opel and colleagues also found childhood maltreatment modified polygenic susceptibility to BMI but in the opposite direction, with exposure to maltreatment reducing the effect of polygenic susceptibility on BMI.30 These opposing findings could be due to different definitions of maltreatment (binary versus continuous score), or different study populations (depressive patients versus population cohort).
We expected to detect effect modification for all outcomes given (a) the association between PGSs and respective phenotypes and (b) the positive associations between childhood maltreatment and CVD subtypes present in our study sample (see Table S4 and interpreting interactions section below). It is difficult to determine whether the lack of effect modification for most outcomes is truly null, underpowered, or due to other factors such as the definition used for risk factors. We have used drinks per week to measure alcohol intake, while previous studies have used units of alcohol per week.31 For drinking and smoking, misclassification bias may occur, which may be differential misclassification if this differs between participants with and without exposure to maltreatment. Integrating the five types of childhood maltreatment into a score, while relevant in the light of dose-response relationships between childhood maltreatment and CVD, may miss maltreatment type-specific effects.
Reports of sex differences in the association between childhood maltreatment subtypes and CVD have been mixed,6,13,32–35 however we did not identify any sex-specific differences. Replication and validation using other cohorts may elucidate whether we missed modification and sex-specific effects due to lack of statistical power.
Positive associations between childhood maltreatment and CVD subtypes and BMI have been reported extensively 5,36 but associations with other CVD risk factors such as blood pressure and lipids are less established.36 We replicated previously observed associations between maltreatment increasing BMI 32–34, smoking behaviour 6,37, and decreasing SBP.38,39
Interpretability of interactions and effect modifications
Interaction and effect modification are terms commonly used interchangeably despite representing subtly different concepts. The underlying statistical model is identical.40
In an interaction, the effects of two causal and independent risk factors are studied, to assess whether their joint exposure leads to a different outcome compared with their independent associations. In effect modification, the effect of one exposure differs (or is modified) according to the value of a second exposure. These exposures are not necessarily causal.18 Both interaction and effect modification can be estimated on the additive or the multiplicative scale. If two exposures are associated with the outcome, an interaction should be detected on at least one of the scales given sufficient statistical power.
In the case of childhood maltreatment, although consistent associations with CVD exist, causality is difficult to assess and therefore the term effect modification better reflects the model considered here. We investigated whether the effect of polygenic susceptibility to a certain trait differed according to values of the childhood maltreatment score. Previous studies have typically only tested the additive scale for interaction.15,16,19,30 Because effect modification can occur on either the additive scale or the multiplicative scale, we analysed both scales.18,19
Relevance for public health
The size of the interaction terms of the few effect modifications detected in this study were very small, even for models employing PGSs with SNPs included using a less stringent significance thresholds. This implies that the clinical impact of these effect modifications is very low, suggesting gene*environment effect modifications are not a major contributor to the excess burden of cardiovascular disease experienced by people who have experienced child maltreatment.
Strengths and limitations
The major strength of our study is the use of a single large cohort, the UK Biobank, giving a large sample size and homogenous measures for our variables. While many studies rely on self-reported measures of phenotypes or disease outcomes, we used hospital and mortality records for CVD and calibrated physical measurements to define phenotypes. Reverse causality can bias analyses, whereby the diagnosis of CVD may lead to medical treatment or altered lifestyle behaviours altering the relative importance of the genotype on disease risk. We therefore excluded all participants with prevalent cases of CVD.
Another strength of this study lies in the use of PGSs rather than a single candidate gene approach, where candidate gene interactions have been shown to be spurious.41 We also repeated analyses using PGSs generated at different significance thresholds, as PGSs including variants detected at less stringent P value thresholds may be more predictive of disease.42 Finally, PGSs are not subject to the same confounding as environmental exposures because they are determined at conception, though they may be affected by population structure. All models were adjusted for all 40 genetic PCs calculated by the UK Biobank to mitigate this bias.
A number of limitations remain. Participation in the mental health questionnaire, an inclusion criterion for our study, has been associated with lower BMI, higher socioeconomic position, and having no T2DM or depression, compared to the rest of the UK Biobank cohort.43 Estimates may be biased by selection bias, where UK Biobank participants are more likely to be female, of higher socioeconomic position, less likely to be obese, to smoke, to drink alcohol on a frequent basis and to experience CVD compared to the rest of the population.24
Childhood maltreatment was determined using self-reported retrospective measures, which could introduce recall bias. An alternative approach would be verifying self-reports with records from social services or other child protecting entities. However, only a fraction of maltreatment cases are reported 44 and thus many cases might be missed. Another limitation is that despite the adjustment for maternal smoking and number of siblings, these might not be the most comprehensive indicators of childhood SEP. Maternal smoking may not be a robust indicator of SEP for this study’s participants as the social patterning of maternal structure was different in the time these participants were born in comparison to current day. This means residual confounding, by childhood SEP or other unmeasured sources, may have biased our findings.
We applied a Bonferroni correction when considering the effect modification P values to address the risk of type 1 error due to multiple tests. This may be overly conservative as cardiovascular risk factors/diseases are correlated phenotypes. However, considering a less stringent P value threshold of P<0.05, there was still little evidence of an interaction between childhood maltreatment and any outcome other than BMI.
Finally, there are caveats with PGSs. Residual confounding by population structure may remain even after adjusting for 40 genetic PCs 45. While our study only included participants of White British Ancestry and used summary statistics from similar populations, some of the PGSs explained a low proportion of the trait’s variance. More so, our findings have limited generalizability to non-White European populations, as PGSs have been found to have very limited portability across ancestries 46.
The disproportionate scrutinising of a single ancestral background is an impediment in our understanding of CVD aetiology and the clinical application of PGSs. Indeed, 83% of GWASs released between 2005 and 2018 were on individuals of White European Ancestry 47, underscoring the importance of the need for trans-ethnic GWASs and the replication of studies such as ours in other populations.
Conclusion
In this analysis on White British individuals from the UK Biobank, we exposure to childhood maltreatment can moderately exacerbate the effects of a genetic propensity to higher BMI. However, we found little evidence that childhood maltreatment modified the effects of other cardiovascular polygenic scores. Our results suggest gene*environment interactions are not a major contributor to the excess burden of CVD experienced by people who have experienced child maltreatment.
Data Availability
All data produced in the present work are contained in the manuscript and in supplementary file. The code used in this analysis is available at github.com/helenaurquijo/gxe_cvd_maltreatment.
Sources of funding
No funding body has influenced data collection, analysis or its interpretations. This research was conducted using the UK Biobank Resource using application 19278. HU is funded by a PhD studentship from the British Heart Foundation (FS/17/60/33474). HU, ALGS, AF, LDH and ARC all work in a unit that receives core funding from the UK Medical Research Council and University of Bristol (MC_UU_00011/1, MC_UU_00011/4 and MC_UU_00011/6). ARC is additionally supported by the University of Bristol British Heart Foundation Accelerator Award (AA/18/7/34219). ALGS is supported by the study of Dynamic longitudinal exposome trajectories in cardiovascular and metabolic non-communicable diseases (H2020-SC1-2019-Single-Stage-RTD, project ID 874739). AF is supported by the National Institute for Health Research (NIHR) Biomedical Research Centre based at University Hospitals Bristol NHS Foundation and the University of Bristol. LDH is funded by a Career Development Award from the UK Medical Research Council (MR/M020894/1). The views expressed are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health.
Disclosures
The authors declare no competing interests.
Author contributions
HU cleaned and analysed the data, interpreted the results, wrote and revised the manuscript. ALGS and ARC had access to the data and assisted in data preparation. ALGS, AF, LDH and ARC all designed the study, interpreted the results, critically reviewed and revised the manuscript and provided supervision for the project. HU and ARC serve as guarantors of the paper. The corresponding author attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted.
Acknowledgements
Quality Control filtering of the UK Biobank data was conducted by R.Mitchell, G.Hemani, T.Dudding, L.Corbin, S.Harrison, L.Paternoster as described in the published protocol (doi: 10.5523/bris.1ovaau5sxunp2cv8rcy88688v). The MRC IEU UK Biobank GWAS pipeline was developed by B.Elsworth, R.Mitchell, C.Raistrick, L.Paternoster, G.Hemani, T.Gaunt (doi: 10.5523/bris.pnoat8cxo0u52p6ynfaekeigi).
Footnotes
Abstract and Results section updated to correct typo in confidence interval value (line 32 and 199).
Abbreviations and Acronyms
- AF
- Atrial fibrillation
- BMI
- Body mass index
- CHD
- Coronary heart disease
- CVD
- Cardiovascular disease
- GWAS
- Genome-wide association studies
- LDL-C
- Low-density lipoprotein cholesterol
- PC
- Principal components
- PGS
- Polygenic score
- RCT
- Randomised control trials
- SBP
- Systolic blood pressure
- SEP
- Socioeconomic position
- SNP
- Single nucleotide polymorphism
- T2DM
- Type-2 diabetes mellitus