ApoE e2 and aging-related outcomes in 379,000 UK Biobank participants

Biomarkers

In Figure 1, we highlight the associations with biomarkers that reached the Bonferroni-adjusted significance level of 5% and showed a 0.1 SD or larger mean difference between e2e2 or e2e3 and e3e3. Compared to e3e3, e2e2 tended to be more associated than e2e3 with biomarkers associated with CAD. Both had lower mean total cholesterol (-0.78 SD in e2e2, p=1.56×10^-314 versus -0.33 SD in e2e3, p<1×10^-323), LDL cholesterol (-1.12 SD in e2e2, p<1×10^-323 versus -0.43 SD in e2e3, p<1×10^-323), lipoprotein A (-0.40 SD in e2e2, p=4.58×10^-62 versus -0.10 SD in e2e3, p=1.41×10^-61), and apolipoprotein B (-1.99 SD in e2e2, p<1×10^-323 versus -0.56 SD in e2e3, p<1×10^-323), plus higher apolipoprotein A1 (0.13 SD in e2e2, p=8.37×10^-11 versus 0.11 SD in e2e3, p=1.51×10^-104), all associated with lower risks of CAD. e2e2 and e2e3, however, had higher mean triglycerides (0.41 SD in e2e2, p=5.42×10^-92 versus 0.11 SD in e2e3, p=1.04×10^-97), which is associated with higher risks of CAD.

e2e2 and e2e3 were associated with lower albumin and higher direct bilirubin and alkaline phosphatase, but the effect of e2e3 was not as striking as that of e2e2, e.g., 0.22 SD higher in e2e2 (p=5.12×10^-24) versus 0.06 SD higher in e2e3 compared to e3e3 for direct bilirubin (p=3.20×10^-28). Additionally, the mean vitamin level of e2e2 was higher than that of e3e3 by 0.1 SD. e3e3 and other ApoE genotypes shared similar vitamin D levels.

There were also associations with various hematology measures (Figure 2 and Supplementary Table 2), including on reticulocyte numbers, with e2e2 being associated with much larger effects than of e2e3. Compared to e3e3, e2e2 had 0.22 SD lower mean reticulocyte count (p=7.90×10^-27), 0.30 SD lower high light scatter reticulocyte count (p=3.78×10^-47), 0.31 SD lower immature reticulocyte fraction (p=3.07×10^-50), and 0.09 SD higher mean reticulocyte volume (p=8.46×10^-6). e2e2 also had 0.43 SD higher red cell distribution width (RDW) than e3e3 (p=2.16×10), 0.14 SD lower mean corpuscular hemoglobin (p=5.84×10^-12), 0.12 SD lower mean corpuscular volume (p=1.51×10^-9), 0.11 SD higher mean platelet volume (p=4.97×10^-8), and 0.13 SD higher mean sphered cell volume (p=2.12×10^-10).

Disease outcomes

The results for diseases with ~80% power or higher to detect odds ratios approximately 1.2 and 1.2² comparing e2e3, and e2e2 to e3e3 are presented in Figure 3. All the disease association results can be found in Supplementary Table 2. CAD and hypertension were the two significant diseases with e2e2 and/or e2e2 effect compared to e3e3 at the 5% Bonferroni-adjusted level. e2e3 had a lower risk of hypertension than e3e3 (OR=0.94, 95% CI: 0.92 to 0.97, p=7.28×10^-7) but the protective effect was reduced in e2e2 (OR=1.01, 95% CI: 0.92 to 1.10, p=0.911). Similarly, e2e3 heterozygotes were protected from CAD but the association was not seen with e2e2: the odds ratio for CAD was 0.87 (95% CI: 0.83 to 0.90, p=4.92×10^-14) comparing e2e3 to e3e3 and was 1.04 (95% CI: 0.90 to 1.19, p=0.635) comparing e2e2 to e3e3.

We found no associations between e2e3 or e2e2 and renal failure or two kidney function biomarkers, creatinine and cystatin (Supplementary Table 2). Similarly, e2e3 or e2e2 was not associated with AMD and dementia.

Chronic Pain, Cognitive Function, Physical Measures and Mortality

e2e3 or e2e2 was not significantly associated with chronic pain, cognitive measures, and physical measures except body mass index (BMI). The mean BMI was 0.07 SD (95% CI: 0.03 to 0.11, p=0.001) higher in e2e2 than in e3e3 (Figure 4), oppositely associated with e4 (Supplementary Table 2). The hazard ratio of death during follow-up in participants comparing e2e3 to e3e3 was 0.96 (95% CI: 0.91 to 1.01, p=0.116) and that comparing e2e2 to e3e3 was 1.06 (95% CI: 0.87 to 1.30, p=0.532) (Figure 5).

Included samples

To avoid genetic confounding, we analyzed European-descent participants (n=451,367, ~90% of the cohort), identified using genetic principal components analysis in detail in Thompson et al. [26]. One in third-degree or closer pairs were removed, leaving a total of 379,703, where the relatedness was determined based on pairwise kinship coefficients, calculated using genome-wide SNP data by the KING software [27].

Aging-Related outcomes

We classified aging-related outcomes into five categories: 1) biomarkers, 2) diseases and chronic pain, 3) mortality, 4) cognitive function, and 5) physical measures. Survival data were updated to Feb 15, 2018 and disease diagnoses to March 31, 2017. Others were surveyed or measured at recruitment/baseline.

Biomarkers

A panel of biomarkers from blood samples at baseline were collected including hematological measures (e.g., white blood count, red blood cell count, and hemoglobin concentration), prognostic biomarkers (e.g., lipids for vascular disease, sex hormones for cancer), diagnostic biomarkers (e.g., HbA1c for diabetes and rheumatoid factor for arthritis), and biomarkers to characterize phenotypes that are not well assessed (e.g., biomarkers for renal and liver function). The full lists including technical details can be downloaded from the links [28, 29]. Each was transformed by the rank-based inverse normal transformation, followed by the z-transformation to correct distribution skewness and to unify the scale across traits.

Diseases and chronic pain

Disease diagnoses were either self-reported at the baseline assessment and verified by a trained nurse during the verbal interview, or from the hospital admission data (HES, hospital episode statistics, covering the period 1996 to March 31, 2017) or the cancer registry. We combined prevalent and incident cases for the analysis of ApoE genotype associations with likelihood of disease. A complete list of International Classification of Disease tenth revision (ICD-10) diagnosis codes used in this study is included in Supplementary Table 3.

Depression and chronic pain at baseline were assessed by survey questions to identify those with a localized pain for 3 months or longer (knee pain 3+ months, back pain 3+ months, and hip pain 3+ months) and those with depressed mood for several days or more in the past two weeks. Additionally, we derived a 49-item frailty index [30] mostly based on diseases and pains considering 60 and older only (not sensible to the middle aged), and applied log+1 transformation to correct distribution skewness.

Mortality

The death status was determined using the death certificate data, where age at death was calculated by date of death minus date of birth in years. For analytical purpose, we also calculated the survival time to the last follow up, which was Feb 15, 2018, for alive participants then.

Cognitive function

We selected two cognitive function measures from touch screen tests at baseline that covered the majority of participants, i.e., reaction time and visual memory errors. The reaction time was measured as the average time used to correctly identify a match in a symbol match game similar to the snap card game. The visual memory errors was measured as the number of errors that a participant made to complete a pairs matching task where 6 pairs of cards were presented for 3 seconds beforehand. Each was log transformed to correct skewness of the distribution. The visual memory errors were right shifted by 1 before the transformation to avoid infinite values from zero visual memory errors.

Physical measures

In baseline physical measures, we included body mass index (BMI), systolic and diastolic blood pressures, any falls in the last year, heel bone mineral density (BMD), lung function measures of FEV1 (forced expiratory volume in 1 second), FVC (forced vital capacity), and FEV1/FVC ratio, Fried frailty (frail or not frail), skeletal muscle mass index [31], and maximal hand grip strength. Any falls in the last year and some elements to derive the Fried frailty were assessed by survey questions. Other measurements were performed at the assessment centers when participants were recruited.

Heel bone mineral density in grams/cm² was estimated based on the Quantitative Ultrasound Index through the calcaneus. The spirometry test was performed using a Vitalograph Pneumotrac 6800 that analyzed 2-3 blows of participants. The Fried frailty [32] was derived using participants aged 60 and older at baseline where the frailty status was confirmed if three or more of the conditions were met, 1) self-reported weight loss (yes/no, based on a survey question to ask weight change compared to one year ago), 2) exhaustion (yes/no, based on a survey question to ask frequency of feeling tired or having little energy over the past two weeks), 3) self-reported slow walking pace (yes/no, based on a survey question to ask usual walking pace: slow if less than 3 miles per hour), 4) lowest 20% of hand grip strength in the same sex group (yes/no), 5) lowest 20% of physical activity in the same sex group (yes/no), by the short version of International Physical Activity Questionnaire (IPAQ) [33].

The maximal hand grip strength of both hands was measured using a Jamar J00105 hydraulic hand dynamometer. The skeletal muscle mass was measured by the skeletal muscle index (SMI) defined by Janssen et al. [31],

$SMI=SMM\text{/}H{t}^2$

with height (Ht) in meters and the skeletal muscle mass (SMM) defined as

$\begin{array}{l}SMM=(H{t}^2\text{/}R)\times 0.401+(gender\times 3.825)\\ +(age\times (-0.071))+5.102\end{array}$

where the Bioelectrical Impedance Analysis resistance (R) in ohms for the whole body was taken by a Tanita BC418MA body composition analyzer; gender was 1 for men and 0 for women and age was measured in years.

Statistical analysis

Aging-related outcomes including time-to-event (survival), continuous, and binary variables were modelled for associations with ApoE genotypes using Cox regression, linear regression, and logistic regression models. Prior to association analyses, continuous variables were log-transformed (cognitive function measures and 49-item frailty) or transformed by the rank-based inverse normal transformation (biomarkers) to correct distribution skewness and further z-transformed so that magnitude of mean differences by genotype between traits are comparable. Each ApoE genotype (e2e2, e2e3, e3e4, or e4e4) was compared with e3e3, adjusted for age at baseline (outcomes measured at baseline) or age at the last update (survival and disease outcomes), sex, assessment center, genotyping array type, and the first five genetic principal components. We highlighted associations with p-values significant at the Bonferroni-adjusted level (p<0.05/106) for the null hypothesis of no e2e3 or e2e2 effect. All the statistical analyses were performed in R version 3.4.1.

Power analysis

To estimate power for continuous traits, we assumed that the three genotype groups e3e3, e2e3, and e2e2 share the same standard deviation for a trait. Given the sample sizes that we have for the three ApoE genotypes, the power to detect a 0.05 standard deviation (SD) mean difference between e3e3 and e2e3 or a 0.1 SD mean difference between e3e3 and e2e2 using an ANOVA F-test at the Bonferroni-adjusted significance level (p<0.05/106), is presented in Supplementary Table 1. Most continuous traits have over 99% power to detect the above differences, except oestradiol (65%) and rheumatoid factor (27%).

To estimate power for binary outcomes, we assumed a multiplicative e2 effect and we aimed to detect the relative risk of 1.2 comparing e2e3 to e3e3 and (1.2)² comparing e2e2 to e3e3, approximately equivalent to the odds ratios from a logistic regression model as the prevalence is low (<0.1). Given the sample sizes that we have for e3e3, e2e3, and e2e2, the power to reject the null hypothesis (p<0.05/106) that both relative risks are 1, is provided in Supplementary Table 1. We focus on disease outcomes with ~80% or higher power, with melanoma cancer (79% power) and traits with better power referred to as “primary” and the rest are considered “secondary”.

Age at death was the only survival outcome in this study, with death status and age at the last follow-up information. For convenience, we approximately calculated the power using death status only as a binary outcome, which should be similar to that of the survival outcome as the death rate is low. We found in actual data analyses that the hazard ratio of death comparing e2e3 or e2e2 to e3e3 (see the Results section) was very similar to the corresponding odds ratio (results not shown).