Abstract
Introduction Genetic susceptibility to various chronic diseases has been shown to influence heart failure (HF) risk. However, the pathogenic mechanisms underlying these associations, particularly the role of phenotypic leukocyte telomere length (LTL), is unknown. We investigated the shared genetic etiology between chronic diseases, various traits, and HF risk, and whether LTL mediates or modifies these relationships.
Methods We conducted prospective cohort analyses on 404,883 European participants from the UK Biobank, including 9,989 incident HF cases. Multivariable Cox regression was used to estimate associations between HF risk and 24 polygenic risk scores (PRSs) for various diseases or traits previously generated in the UK Biobank using a Bayesian approach. We assessed multiplicative interactions between the PRSs and LTL previously measured in the UK Biobank using quantitative PCR. Mediation analyses were conducted to estimate the proportion of the total effect of PRSs acting indirectly through LTL, an integrative marker of biological aging.
Results We identified 9 PRSs associated with HF risk, including those for various cardiovascular diseases or traits, rheumatoid arthritis (P=1.3E-04), and asthma (P=1.8E-08). Additionally, longer LTL was strongly associated with decreased HF risk (P-trend=1.7E-08). Notably, the asthma PRS had a super-multiplicative interaction with LTL (P-interaction=2.8E-03). However, LTL mediated only 1.13% (P<0.001) of the total effect of the asthma PRS on HF risk.
Conclusions Our findings shed light onto the shared genetic etiology between HF risk, asthma, rheumatoid arthritis, and other traits. Longer LTL strengthened the genetic effect of asthma on HF, supporting their utility in risk stratification analyses.
Introduction
Heart failure (HF) is a phenotypically complex syndrome that is commonly characterized by the diminished ability of the ventricle to pump or fill with blood 1–4. HF is a downstream consequence of most cardiovascular diseases (CVD) including coronary heart disease, high blood pressure, atrial fibrillation (AF), and cardiomyopathies 5. Genome-wide association studies (GWAS) have found a modest contribution of genetic susceptibility to HF risk, with heritability estimates ranging from 8.8% 6 to 26% 7. Thus far, 12 independent genome-wide significant variants have been identified to be associated with HF among European populations 6, 8. All of the identified genetic variants had associations with coronary artery disease (CAD), AF, or reduced left ventricular function, which strongly supports the downstream nature of HF relative to CVD 6, 8. Besides CVD, recent studies have shed light on bi-directional relationships between HF development and various cancers 9–12, which may be attributed in part to shared biologic pathways 13, 14. A prominent molecular pathway for the initiation and progression of both HF and cancers is increased inflammatory states or dysfunction 15, including those due to autoimmune diseases such as rheumatoid arthritis 16, 17. Taken together, these observations suggest shared genetic susceptibility across CVDs, cancers, various inflammatory phenotypes and the pathogenesis of heart failure. However, the biological processes underlying these relationships are unclear. Given that chronological age is a strong risk factor for heart failure (as well as other chronic diseases), genetic susceptibility potentially influences heart failure risk through accelerated biological aging or weathering. One such biomarker reflective of biological aging is leukocyte telomere length (LTL).
Telomeres are hexameric (TTAGGG)n repeats at the ends of linear chromosomes that protect the genome from progressive degradation after each round of mitosis. LTL has a modest genetic heritability of approximately 8.1% 18, 19, with the remaining variation explained largely by potential missing heritability, non-genetic environmental factors, and technical variables. LTL has an apparent dual-nature and reflects two broad components. First, LTL reflects exogenous environmental exposures that cause genotoxic damage to the DNA of hematopoietic stem cells, their derivative leukocytes, as well as highly replicative target tissues 20. Second, LTL reflects endogenous biological pathways related to biological aging or weathering 21–30. Taken together, LTL can be considered as an integrative biomarker of genomic instability that captures the cumulative burden of biological stressors across the lifespan. Shorter phenotypic LTL, which reflects a higher load of biological stressors, has been consistently found to be associated with increased risk of HF 18, 31, CVD 32, and coronary heart disease33. Additionally, measured or genetically-determined LTL have been reported to have relationships of various directions with different cancers 34–40 as well as inflammatory conditions 41. However, the precise role of LTL in the pathogenic mechanism underlying the relationship between genetic susceptibility and HF risk is unknown.
To investigate these gaps in knowledge, we leveraged the extensive questionnaire, clinical, and molecular/genetic data from the UK Biobank, a prospective cohort study of nearly half a million adults residing in the United Kingdom. The UK Biobank is among the premier cohort studies for the investigation of genetic etiology among European populations and is the single largest resource for human telomere length data in the world 18, 19. Our overarching goal was to gain mechanistic insight into the role of phenotypic LTL in influencing the effect of genetic susceptibility on HF risk. Specifically, we were interested in: 1) agnostically exploring the effect of genetic susceptibility to cardiovascular diseases or traits, cancers, metabolic diseases, and inflammatory conditions on HF risk; 2) investigating whether LTL, particularly the environmentally-determined components, modifies the strength of associations between genetic susceptibility and HF risk; and 3) estimating the degree in which LTL, particularly the biological aging component, mediates the total effect of genetic susceptibility on HF risk. Our findings improve the understanding of the shared biological mechanisms underlying HF and other non-malignant and malignant diseases, as well as support the use of LTL and genetics in risk stratification analyses to identify high-risk subgroups that could benefit from early intervention.
Materials and Methods
The UK Biobank dataset used for this study is publicly available at: https://bbams.ndph.ox.ac.uk/ams/. Our analyses were conducted under UK Biobank project number: 28072 (datasets from August 2022 – June 2023). Additional metadata and codes are available from the corresponding author upon reasonable request. The detailed methods are available within the Online Supplement. Briefly, we used a multi-step approach to investigate the interrelationships between genetic susceptibility to various diseases or traits, LTL, and HF risk (Figure 1). Among participants of European ancestry, we first agnostically assessed the prospective associations between 24 previously derived CVD-, cancer-, and inflammation-related polygenic risk scores (PRSs) 42 and future risk of developing HF. Second, among the PRSs identified to be significantly associated with HF risk, we then investigated whether the PRSs statistically interact with measured LTL to influence HF risk. Among PRSs with evidence of multiplicative interaction with LTL, we conducted further analyses stratified by quartiles of LTL. In parallel, we agnostically assessed associations between the 24 PRSs and LTL measured at the time of study enrollment. Here, we qualitatively looked for overlapping findings with the PRS-HF analyses. Lastly, we determined whether the effect of the significant PRSs on HF risk was mediated through biological aging processes reflected by altered LTL.
Results
Dose-response relationships between polygenic risk scores for various diseases or traits and risk of heart failure
Among 404,883 participants of European ancestry, we identified 9,989 incident cases of HF (first reported in-patient hospitalization) throughout the 12.3 (1.8 SD) year follow-up. Among the HF cases, the average follow-up time to HF diagnosis was 8.3 (3.3 SD) years and the average age at diagnosis was 70.3 (7.0 SD) years.
First, we assessed the associations between 24 CVD-, cancer-, metabolic-, and inflammation-related polygenic risk scores (PRSs) 42 and future risk of developing HF (Figure 1) using multivariable Cox regression models adjusted for potential confounders as well as four genetic principal components to control for population stratification. The genome-wide data used to generate the PRSs were from external GWAS independent of the UK Biobank. We identified 9 PRSs that were significantly associated with HF risk that are reflective of increased genetic susceptibility to various cardiovascular diseases and traits, a respiratory disease, adiposity, and an autoimmune disease (Table 1). Notably, we found that increased genetic susceptibility to asthma, a prominent respiratory and inflammatory disease, was significantly associated with increased HF risk (P=1.8E-08; Table 1).
As expected, among the PRSs for cardiovascular diseases and traits, increased genetic susceptibility to AF (P<1.0E-28), overall CVD (P<1.0E-28), coronary artery disease (CAD; P<1.0E-28), hypertension (HT; P<1.0E-28), and ischemic stroke (ISS; P<1.0E-28) were significantly associated with elevated risk of HF in a dose-dependent manner (Table 1). However, the positive association between genetic susceptibility to venous thromboembolism (VTE) and HF risk was not significant after correction for multiple comparisons (P=0.01).
When examining autoimmune disorders, we found a positive dose-response relationship between genetic susceptibility to rheumatoid arthritis (RA) and HF risk (P=1.3E-04). However, associations with Crohn’s Disease (CD) were not significant after Bonferroni correction for multiple comparisons (P=0.04; Table 1). With respect to adiposity, genetically-predicted body mass index (BMI) was positively associated with HF risk (P=2.6E-06). When examining circulating cholesterol, increased genetically-predicted high-density lipoprotein (HDL) levels were significantly associated with lower risk of HF (P=7.9E-04), while increased low-density lipoprotein (LDL) levels were non-significantly associated with higher HF risk (P=0.03; Table 1). With respect to malignant diseases and HF risk, we did not detect any significant associations for genetic susceptibility to colorectal cancer (CRC), breast cancer (BC) and ovarian cancer (EOC) among women, and prostate cancer (PC) among men (Table 1).
LTL modifies the effect of asthma genetic susceptibility on HF risk
Longer measured LTL, which reflects less exposure to environmental stressors, was significantly associated with decreased risk of HF (HRper 1 SD LTL=0.94, 95% CI: 0.92-0.96, P-trend=1.7E-08) independent of the PRSs. We then considered the 9 PRSs that were previously identified to be significantly associated with HF risk (Table 1) and tested for multiplicative interaction with measured LTL (Table 2). We found evidence for a super-multiplicative interaction between measured LTL and the PRS for asthma susceptibility, with longer LTL strengthening the positive asthma-HF association (P-interaction=2.8E-03). Additional Cox regression analyses stratified by quartiles of measured LTL supported a similar strengthening of the asthma-HF association with progressively longer LTL (Figure 2). However, we did not detect evidence for multiplicative interactions between LTL and the other PRSs (Table 2).
The effect of genetic susceptibility to various disease or traits on LTL measured at enrollment
We assessed whether the 24 PRSs were associated with altered LTL measured at enrollment and identified 13 significant findings (Table 3). We found that increased genetic susceptibility to asthma (P=1.7E-06), overall CVD (P=3.0E-18), celiac disease (CED; P<1.0E-28), CAD (P= 6.5E-13), ISS (P=1.2E-06), systemic lupus erythematosus (SLE; P= 7.2E-15), type 1 diabetes (T1D; P= 8.8E-10), and type 2 diabetes (T2D; P= 1.8E-06) as well as genetically predicted HbA1C levels (P= 1.2E-04) were associated with shorter LTL, which reflects increased biological aging (Table 3). Additionally, we found that higher genetic susceptibility to EOC (P= 3.8E-15), melanoma (MEL; P< 1.0E-28), and PC (P= 6.8E-13) as well as genetically-predicted LDL levels (6.2E-06) were associated with longer LTL (Table 3). Taken together, these findings suggest shared genetic etiology between these diseases or traits with the biological aging component of LTL (Table 3). Among these 13 significant findings, the PRSs for asthma, CVD, CAD, and ISS overlapped with the significant associations identified in the PRS-HF analyses (Table 1), which suggests potential interrelationships between genetic susceptibility to cardiovascular and pulmonary outcomes, phenotypic LTL, and future risk of HF. These findings prompted us to assess whether the effects of the 9 previously identified PRSs on HF risk (Table 1) were potentially mediated through altered LTL (Figure 1).
Limited evidence for LTL as a mediator of PRS-HF associations
We conducted formal mediation analyses, which decomposes the total effect of each of the PRSs on HF risk into four components: 1) the pure direct effect; 2) only interaction, but not mediation; 3) both mediation and interaction; and 4) only mediation, but not interaction (i.e., pure indirect effect) 43. To facilitate the mediation analyses, we dichotomized each of the 9 PRSs associated with HF risk by their respective 75th percentiles into a “high” genetic susceptibility category and a “normal-to-low” genetic susceptibility category (referent). Among the 9 PRSs, we found 8 that were associated with HF with statistically significant total excess relative risks (Table 4). The 8 noteworthy PRSs included those related to asthma, AF, BMI, CVD, CAD, HDL, HT, and ISS. Among these 8 PRSs, we were primarily interested in the excess relative risk due to the pure indirect effect through LTL (i.e., the mediated effect). The strongest mediated effect through LTL was for the asthma PRS, which accounted for a very small but statistically significant 1.13% of the total effect (P<0.001; Table 4). The mediated effect through LTL for the BMI, CVD, CAD, HT, and ISS PRSs were also statistically significant (P<0.001) but accounted for only 0.23% - 0.60% of their respective total effects (Table 4).
Discussion
Shared genetic etiology between various chronic diseases or traits and HF risk
We conducted comprehensive analyses to investigate the shared genetic etiology between various chronic diseases or traits and HF risk in a large-scale prospective cohort study of Europeans. First, we found that increased genetic susceptibility to cardiovascular, respiratory, autoimmune, and adiposity related diseases or traits was associated with future risk of HF in a population without clinically diagnosed CVD at enrollment. These findings suggest shared genetic etiology between these diseases or traits and the pathogenesis of HF. The positive findings for CVD are consistent with the fact that HF is a terminal manifestation of most forms of CVD. The findings for RA were of particular interest because of the growing appreciation of inflammatory or autoimmune disorders in HF pathogenesis 44, while asthma was of interest because of the established link between pulmonary diseases, cardiac function, and inflammation.
The interrelationships between asthma genetic susceptibility, LTL, and HF risk
To gain additional mechanistic insight, we also investigated how LTL influences the effect of genetic susceptibility on HF risk. In agreement with previous analyses 18, 31, we found that longer phenotypic LTL, which reflects decreased environmental stressors and biological aging, was associated with lower risk of HF. Subsequently, we found evidence that longer phenotypic LTL acted as an effect modifier that strengthened the positive association between asthma genetic susceptibility and HF risk. This effect modification was likely attributed to the environmentally-determined components reflected by LTL, which explain most of the variance in telomere length. These findings support the potential utility of incorporating LTL and genetic data into future risk stratification analyses. However, we found negligible evidence for indirect effects mediated through the biological components reflected by LTL. As such, the environmental components reflected by LTL may be of greater etiologic importance to HF compared to the biological or genetic components.
The genetic link between asthma and HF risk
Asthma is a chronic respiratory and inflammatory disease that is characterized by obstructed airflow and bronchospasms 45. Previous studies have found some evidence linking asthma to HF 46 and its risk factors 47. A cohort study conducted in Northern California found that those with asthma had over double the risk of HF 48. Additionally, a Mendelian Randomization (MR) analysis conducted in a European population from FinnGen and the Heart Failure Molecular Epidemiology for Therapeutic targets consortium found evidence suggesting a causal association between asthma and HF risk 49. Our PRS findings confirm the previously conducted MR analyses 49 and expand upon the molecular mechanism by elucidating the role of LTL as an effect modifier.
Even though phenotypic LTL was inversely associated with HF independent of genetic susceptibility, it had a super-multiplicative interaction with the asthma PRS to strengthen the association with HF. The reason for the direction of this statistical interaction is unclear. Given that phenotypic LTL reflects endogenous biological processes related to aging as well as cumulative burden of exogenous environmental exposures, it is possible that each genetic variant used to derive the PRS could be interacting differently with the various components reflected by LTL. Complementary functional experiments are needed to further elucidate the nature of any potential biological interaction.
Genetic liability to other non-malignant and malignant diseases and their role in altering LTL
In addition to the main results, we also found that genetic susceptibility to various cardiovascular diseases and traits as well as cancers was associated with altered phenotypic LTL measured at the time of enrollment. The observed relationships between cardiovascular disease and traits and shorter LTL were consistent with previous prospective and retrospective studies 33, 50 and further support the role of their shared genetics 51, 52. Interestingly, we also found that genetic susceptibility to melanoma, ovarian epithelial cancer, and prostate cancer was associated with longer phenotypic LTL, which was in concordance with previous MR analyses in European and Asian populations 34. However, the relationship between telomere length and cancer risk is complex and nuanced. In prospective studies, LTL associations were equivocal for risk of breast cancer and other sites, 35, 37, 53, 54; shorter LTL was associated with increased risk of bladder 55, esophageal 56 and overall cancer 57; and longer LTL was associated with increased risk of non-Hodgkin lymphoma 58 and potentially melanoma 59. Most notably, longer phenotypic and genetically-determined LTL has been consistently found to be associated with increased risk of lung adenocarcinoma among Europeans and East Asians 34, 38–40, 60–63. These positive associations defy the traditional expectation of telomere dynamics in cancer development and the precise biological mechanism underlying these paradoxical relationships is unclear. A potential explanation is that longer telomeres in target tissue (indirectly reflected by LTL) may lead to delayed senescence of pre-cancerous cells, which facilitates the accumulation of somatic abnormalities that potentially drive carcinogenesis and cancer progression 40, 60,20.
Strengths and limitations
Our study had numerous strengths. First, the UK Biobank is the world’s largest prospective cohort study to date that measured LTL in pre-diagnostic biospecimens. The sample size and rich data optimized the statistical power to detect complex associations, which is particularly important when investigating statistical interactions. Furthermore, the prospective study design allowed temporality to be established between genetic susceptibility, phenotypic LTL, and HF risk, which permitted mediation analyses to evaluate the precise interrelationships between these variables. Additionally, the PRSs used in our analyses were derived for the UK Biobank using a Bayesian method with genetic variants across the entire genome, as opposed to conventional PRSs which are comprised of a limited number of genome-wide significant variants 42. Lastly, the UK Biobank was linked to national hospital registries, which likely captured most HF cases in the cohort, thus reducing the potential for outcome misclassification.
Our study had some limitations. Most of the UK Biobank participants were of White European ancestry, which limited the generalizability of our findings. The observed interrelationships could be potentially different among ethnic subgroups with different genomic architecture and exposure profiles that influence LTL and HF risk. Additionally, we had limited information on heart failure subtypes, severity, and whether left ventricle ejection fraction was preserved or not. Lastly, phenotypic LTL measured using qPCR has become a mainstay among epidemiologic studies of chronic disease. However, the adjusted relative telomere-to-single copy gene (T/S ratio) reflects the average telomere length in a population of white blood cells but does not provide information on the absolute telomere length of specific chromosomal arms.
In summary, our findings strongly suggest that people with both longer LTL and increased genetic liability to asthma may have elevated risk of HF, a highly fatal consequence of cardiac dysfunction that has a mortality rate comparable to some cancers 64,65. These findings contribute to the understanding of the shared molecular/genetic mechanisms underlying HF and non-malignant respiratory diseases, as well as the role of LTL as an effect modifier in the underlying mechanisms. Importantly, we potentially identified higher-risk subpopulations for HF, which warrants further investigation of LTL and genetics in future risk stratification analyses.
Data Availability
Data used for this study are publicly available at: https://bbams.ndph.ox.ac.uk/ams/ Meta-data and codes produced in the present study are available upon reasonable request to the authors.
CRediT author statement
Jason Y.Y. Wong-Conceptualization, methodology, formal analysis, data curation, writing original draft
Batel Blechter-Methodology, formal analysis, writing review and editing
Zhonghua Liu-Methodology, writing review and editing
Jianxin Shi-Methodology, formal analysis, writing review and editing
Véronique L. Roger-Supervision, conceptualization, writing original draft
Online Supplement
Methods
Study design
The UK Biobank is a prospective cohort study with extensive biospecimen collection at the time of enrollment. The study design, rationale, and methodologies have been described in detail 66, 67. In brief, the source population was adults aged 40-69 years who lived ≤40 km of 22 study assessment centers throughout the UK. Approximately 9.2 million people registered in the UK’s National Health Service were mailed study invitations, whereas a subset of 503,317 people (5.5%) visited the assessment centers during the recruitment period of 2006-2010 66. Volunteer participants who provided electronic informed consent were given touchscreen questionnaires, physical examinations, and provided biospecimens for biomarker analyses. As part of UK Biobank project number: 28072, our dataset obtained in August 2022 included a total of 502,409 participants. The UK Biobank is a continuously updated dataset without periodic data freezes.
Heart failure diagnosis
Heart failure was defined using in-patient hospital diagnoses coded according to the International Classification of Disease version 9 (ICD-9) and 10 (ICD-10) classifications. ICD-10 codes I50.0, I50.1, I50.9, I11.0, I13.0, and I13.2 along with ICD-9 codes 428.0 and 428.1 were used to define heart failure.
Prospective follow-up
For each participant, follow-up time started at the date of visit to the assessment center in 2006-2010 and ended at the date of first incident heart failure diagnosis (outcome), death (censored), or administrative censoring (i.e., September 20th, 2021, for England and Wales and October 31st, 2021, for Scotland), whichever came first. Vital status as well as date and primary underlying cause of death was provided by the NHS Information Centre and the NHS Central Register Scotland. The UK Biobank study was approved by the National Information Governance Board for Health and Social Care and the NHS North West Multicenter Research Ethics Committee. All volunteer participants provided electronic informed consent.
Measured phenotypic leukocyte telomere length (LTL)
DNA extraction, LTL measurement, and quality control were performed as previously described 68. Data from the multiplex quantitative polymerase chain reaction (qPCR) assay were used to calculate the ratio of telomere repeat copy number (T) to single-copy gene copy number (S; beta-globin) for each subject. This ‘T/S ratio’ for each subject was then divided by the ‘T/S ratio’ of a calibrator sample included on every run to calculate the ‘relative T/S ratio’, which reflects the average telomere abundance across all chromosomes in leukocytes of an individual. The ‘adjusted relative T/S ratio’ accounted for batch variation over time 68.
Polygenic Risk Scores (PRSs) for various diseases and quantitative traits
The 24 PRSs (a subset of 28 diseases and 8 quantitative traits) used in our analyses were previously derived as part of UK Biobank project number 9659 as described in detail (https://biobank.ndph.ox.ac.uk/ukb/label.cgi?id=301) 42. Briefly, genotype data from 825,927 genetic variants were generated using a custom Axiom genotyping array, followed by genome-wide imputation. The PRS algorithms were constructed from trait-specific meta-analyses using a Bayesian approach combining data across multiple ancestries and related traits where appropriate 42. The PRS values for each participant were calculated as the genome-wide sum of the per-variant posterior effect size multiplied by allele dosage 42. A principal component (PC)-based ancestry centering step 69 was used to center the PRS distributions across all ancestries. PRS distributions were also standardized to have approximately unit variance within ancestry groups.
The Bayesian PRSs were generated by Genomics plc (https://www.genomicsplc.com/) in partnership with Our Future Health (https://ourfuturehealth.org.uk/). The performance metrics and validation of these PRSs were as described elsewhere 42. Additional meta data, datasets, and code can be found at: https://github.com/Genomicsplc/ukb-pret/blob/main/UK_Biobank_PRS_Release_Dataset_Notes.md, https://biobank.ndph.ox.ac.uk/showcase/refer.cgi?id=5202, and https://zenodo.org/record/6631952. The genetic data used to derive the standard PRSs were from external GWAS data independent of the UK Biobank. The standard PRSs were generated for all UK Biobank participants. The ‘enhanced’ PRSs previously derived by Thompson et al. 42, which incorporates UK Biobank and external data, were not analyzed in our study because they were only available for a subset of the participants and undergoing further development.
Statistical analyses
Associations between PRSs and HF risk
Multivariable Cox regression models were used to estimate the hazard ratios (HRs) and 95% confidence intervals (CIs) of incident diagnosed heart failure (first reported in-patient hospital admission), in relation to the per 1 SD increase of each of the 24 continuous PRSs in separate analyses. The models were further adjusted for potential confounders including study assessment center, age at entry, body mass index (BMI), sex, smoking status (never, former, current), alcohol consumption, diabetes status, material deprivation (continuous Townsend deprivation index), HbA1C levels, hypertension status, and four genetic principal components. For breast cancer and epithelial ovarian cancer PRSs, we restricted the analyses to women and removed sex from the models (4137 cases; 231,376 participants). For the prostate cancer PRS, we restricted the analyses to men and removed sex from the models (6,169 cases; 186,345 participants). Follow-up time was used as the underlying timescale to improve consistency with previous analyses 70, 71. Evidence for multiplicative effect modification was assessed using cross-product terms in the models between the PRSs and continuous natural log-transformed LTL. We conducted additional PRS-HF analyses stratified by quartiles of LTL. Proportional hazards assumptions were evaluated using Supremum tests and visually examining plots of Schoenfeld residuals over follow-up time. Bonferroni correction of the alpha significance threshold (i.e., 0.05 / number of tests) was used to account for multiple comparisons.
Associations between PRSs and LTL
Multivariable linear regression models were used to separately estimate associations between each continuous PRS and continuous natural log-transformed LTL measured at enrollment. The models were adjusted for age at entry, study center, BMI, sex, smoking status (never, former, current) alcohol intake, material deprivation, four genetic principal components, and leukocyte differentials.
Mediation analyses
We evaluated whether the association between the noteworthy PRSs and HF risk was mediated through continuous LTL using the Med4Way package in STATA/MP 13.0 (StataCorp LP, College Station, TX) 43. This approach uses the delta method to decompose the total effect of exposure on outcome into four components that correspond to the portion of the total effect that is due to: 1) pure direct effect of exposure on outcome; 2) interaction but not mediation; 3) both mediation and interaction ; and 4) pure indirect mediated effect 43. To facilitate the mediation analyses, we dichotomized each of the noteworthy PRSs by their respective 75th percentiles into a ‘high’ genetic susceptibility category and a ‘normal-to-low’ genetic susceptibility category (referent). The Cox regression models for the outcome and linear regression models for the mediator were minimally adjusted for age at entry, sex, study assessment center, material deprivation, smoking status (never, former, current), BMI, alcohol intake, hypertension status, diabetes status, and four genetic principal components. LTL was natural log-transformed to approximate a normal distribution.
Acknowledgements
This study was supported by intramural funding from the National Cancer Institute and National Heart, Lung, and Blood Institute. We thank Drs. Qing Lan and Nathaniel Rothman for their continuing support. Additionally, we extend our deepest appreciation to Lisa Finkelstein, Jillian Varonin, and the UK Biobank Access Team for helping us navigate complex international data agreements. We declare no conflicts of interest.