Integrating the environmental and genetic architectures of aging and mortality

Abstract

Both environmental exposures and genetics are known to play important roles in shaping human aging. Here we aimed to quantify the relative contributions of environment (referred to as the exposome) and genetics to aging and premature mortality. To systematically identify environmental exposures associated with aging in the UK Biobank, we first conducted an exposome-wide analysis of all-cause mortality (n = 492,567) and then assessed the associations of these exposures with a proteomic age clock (n = 45,441), identifying 25 independent exposures associated with mortality and proteomic aging. These exposures were also associated with incident age-related multimorbidity, aging biomarkers and major disease risk factors. Compared with information on age and sex, polygenic risk scores for 22 major diseases explained less than 2 percentage points of additional mortality variation, whereas the exposome explained an additional 17 percentage points. Polygenic risk explained a greater proportion of variation (10.3-26.2%) compared with the exposome for incidence of dementias and breast, prostate and colorectal cancers, whereas the exposome explained a greater proportion of variation (5.5-49.4%) compared with polygenic risk for incidence of diseases of the lung, heart and liver. Our findings provide a comprehensive map of the contributions of environment and genetics to mortality and incidence of common age-related diseases, suggesting that the exposome shapes distinct patterns of disease and mortality risk, irrespective of polygenic disease risk.

Fig. 1. Study overview.

a, After participant exclusions, UKB participants were split into independent discovery, replication and validation sets. Missing values were imputed separately within each group using random forest multiple imputation, resulting in five imputed datasets for each dataset. b, Among UKB participants recruited in England (n = 436,891), an exposome-wide association study (XWAS) for all-cause mortality was conducted using the discovery and replication sets. The discovery and replication sets were then pooled, and further analyses were conducted in the full sample to identify and remove replicated exposures that were sensitive to reverse causation (disease sensitivity) and mismeasurement (PheWAS per exposure). The remaining exposures were then tested cross-sectionally for associations with a previously developed proteomic aging clock (n = 45,441). We then conducted a final sensitivity analysis in the participants recruited in England (n = 436,891) to remove exposures sensitive to correlation bias (cluster analysis). c, Exposures surviving all analyses in b were then tested in relation to 25 age-related biomarkers, 25 age-related diseases and 3 common disease risk factors (hypertension, obesity and dyslipidemia). For mortality and each age-related disease, the relative contributions of age and sex, polygenic risk and exposome were calculated via multivariable Cox proportional hazard models. Multivariable models were validated in participants recruited in Scotland or Wales (n = 55,676), who were held out from all other analyses. Figure created with BioRender.com.

Fig. 2. Environmental architecture of mortality in the UKB.

a, The correlation (Pearson r) between regression coefficients (beta) for the association between each exposure and mortality calculated separately in women (n = 237,637) and men (n = 199,257). The P value for the significance of the Pearson correlation is also given. b, Volcano plot of log-transformed P values and fold change (calculated as log₂ of the HR) for all XWAS associations for mortality in the final pooled analysis. Each point represents the effect and P value for the association between a single exposure and all-cause mortality from a Cox proportional hazard model in the XWAS discovery analysis (n = 218,446). Exposures that were FDR significant in both the discovery and replication stages are colored, whereas associations that were not replicated are indicated in dark gray and grouped in the category *nonreplicated. The top 20 points according to P value are labeled. c, A heat map of β coefficients representing associations between all exposures (only those passing disease and phenome-wide sensitivity analyses) and mortality (from the XWAS discovery analysis; n = 218,446) and proteomic aging (n = 45,441). d, Importance of individual exposures, as assessed by a multivariable model including age, sex and all 25 exposures associated with mortality and proteomic aging that passed all sensitivity analyses (n = 436,891). The importance of each variable was determined using a Wald test from ANOVA, and was calculated as the proportion of that variable’s Wald Χ² relative to the total model Χ². Note that the y-axis values were transformed by taking the square root to improve visualization. Physical activity was measured using the International Physical Activity Questionnaire (IPAQ). LTPA, leisure time physical activity; OPA, occupational physical activity; PM, particulate matter.

Fig. 3. Forest plot of exposome associations with all-cause mortality (n = 436,891) in multivariable models for each individual cluster of correlated exposures.

The models were Cox proportional hazard models calculated using age as the timescale, stratified by 5-year birth cohorts and sex, and adjusted for UKB assessment center, years of education, household income and ethnicity (only if the covariate was not already in the cluster model). The regression estimates are shown with 95% confidence intervals, and estimates not significant at P < 0.05 are shown as hollow points. The P values were not adjusted for multiple comparisons. Physical activity was measured via the IPAQ.

Fig. 4. Environmental architectures of age-related biological mechanisms and diseases in the UKB.

a, A heat map showing associations between each mortality-associated exposure and aging biomarkers. b, A heat map showing associations between each mortality-associated exposure and common disease risk factors. c, A heat map showing associations between each mortality-associated exposure and proteomic aging. d, A heat map showing associations between each mortality-associated exposure and age-related chronic diseases. The colors in the heat maps represent regression coefficients (β) for associations between exposures and biomarkers/diseases. A line annotation track is shown that counts the total number of FDR significant associations for each outcome. For the heat map in a, an additional annotation track shows the primary biological mechanism associated with each aging biomarker. For nominal categorical variables with more than one response level, the association for the level with the strongest P value is reported and the exposure’s label reflects the response category shown. COPD, chronic obstructive pulmonary disease; FDR, false discovery rate; HDL, high-density lipoprotein; IGF-1, insulin-like growth factor-1; LDL, low-density lipoprotein.

Fig. 5. Combined environmental and genetic architectures of mortality and age-related diseases.

a, A plot showing R² values calculated across studied outcomes for several sequential multivariable models: model 1 containing age and sex (purple); model 2 containing age, sex and PRS (yellow); and model 4 containing age, sex, PRS and exposome (green). If a PRS was not available for a particular outcome, then the green R² shows the results from model 3 (age, sex and exposome). The R² values are shown from the validation analyses (n = 55,676). b, The variable importance for age, sex, polygenic risk and exposome for all outcomes studied in model 4 conducted among UKB participants recruited in England (n = 436,891). The importance of each variable was determined using a Wald test from ANOVA, and was calculated as the proportion of that variable’s Wald Χ² relative to the total model Χ² for each category so that they sum to 1. The total importance for PRS also includes the genetic principal components and genotyping batch covariates used. PRS used for mortality models include PRS for all other diseases and phenotypes shown (22 in total). Note that PRS information was not available for liver cancer or lymphoma and is not included in the models. Ovarian, breast and prostate cancer models were sex specific and sex was not included in model 4 for these outcomes. AD, Alzheimer’s disease; COPD, chronic obstructive pulmonary disease; PRS, polygenic risk score.

Extended Data Fig. 1. Mortality and disease incidence rates among UK Biobank participants.

(a) The number of deaths in females and males according to age at death (in years) among UK Biobank participants who died during follow up (n = 31,716). (b) Numbers of prevalent and incident cases for all age-related diseases studied among UK Biobank participants recruited in England (n = 436,891). Note that diseases are put into two groups with different x-axis scales, since some diseases had far more cases than others.

Extended Data Fig. 2. Mortality XWAS associations by different intervals of follow up time.

Correlation between mortality XWAS regression estimates (betas) calculated in the full pooled sample (x-axis; n = 436,891) and the subset of participants excluding those who died within the first 4 years of follow up (y-axis; n = 431,394). Correlation between betas (Pearson r) is shown, as is the p-value for the correlation. A best fit line is fitted by regressing the betas from the y-axis onto the betas from the x-axis.

Extended Data Fig. 3. Mortality XWAS associations accounting for prevalent disease.

Correlation between mortality XWAS regression estimates (betas) calculated in the full analytic sample (x-axis; n = 436,891) and the subset of participants with no disease or poor health at baseline (y-axis; n = 221,067). Correlation between betas (Pearson r) is shown, as is the p-value for the correlation. A best fit line is fitted by regressing the betas from the y-axis onto the betas from the x-axis. Labeled points are those variables that were flagged during the disease indicator interaction analysis.