Plasma proteins, cognitive decline, and 20‐year risk of dementia in the Whitehall II and Atherosclerosis Risk in Communities studies

Abstract Introduction Plasma proteins affect biological processes and are common drug targets but their role in the development of Alzheimer's disease and related dementias remains unclear. We examined associations between 4953 plasma proteins and cognitive decline and risk of dementia in two cohort studies with 20‐year follow‐ups. Methods In the Whitehall II prospective cohort study proteins were measured using SOMAscan technology. Cognitive performance was tested five times over 20 years. Linkage to electronic health records identified incident dementia. The results were replicated in the Atherosclerosis Risk in Communities (ARIC) study. Results Fifteen non‐amyloid/non‐tau–related proteins were associated with cognitive decline and dementia, were consistently identified in both cohorts, and were not explained by known dementia risk factors. Levels of six of the proteins are modifiable by currently approved medications for other conditions. Discussion This study identified several plasma proteins in dementia‐free people that are associated with long‐term risk of cognitive decline and dementia.

Based on these measures, we created a global cognitive score by first standardizing the distribution of each test domain measured in follow-up visits to the baseline score to create z-scores with mean 0 and SD 1. We then summed the domain specific scores at each phase and standardized the summary score to the baseline summary score; this approach minimizes measurement error inherent in individual tests. 29 After dementia diagnosis, participants were rarely able to complete the cognitive tests and for this reason, their global test score was set to -3SD at the phase following the dementia diagnosis. Based on the global scores we derived a cognitive decline slope for each participant and used this as the outcome.

Imputation model
In the Whitehall II study no participants were missing protein measurements but some covariates were missing; 26% had missing education status, 11% were missing APOE and 10% BMI. Education variable was introduced to questionnaire during data collection meaning that it is likely to be missing at random. APOE values were measured in a random sample, suggesting values are also missing at random. BMI values are missing due to lack of funds to cover sickness absences of the research nurse collecting height and weight at phase 5 and are also likely to be missing at random. Percentages of participants missing data for other variables were very low and unlikely to affect the imputation results regardless of the missing data mechanism. For these reasons, we used the missing at random assumption and conducted multiple imputation for covariate-adjusted analyses. Our imputation model included age, sex, education, socioeconomic status, systolic blood pressure, total cholesterol, high-density lipoprotein cholesterol, antihypertensive medication, diabetes status, ApoE status, alcohol consumption, General Health Questionnaire, BMI, and smoking status. Continuous variables were imputed with linear regression, binary with logistic regression, categorical with ordered logistic regression, and right tailed integer data with negative binomial regression. Based on percentage of missing data and substantive model parameters from 100 iterations we chose to use 5 iterations and 30 imputations in the final analyses. We used an imputation model based on substantive-model compatible fully conditional specification, which may be a better approach in survival analyses and in the presence of competing risks than the ordinary fully conditional model. 130 The two-step approach used to identify protein-dementia association for replication We used a two-step approach in Whitehall II to identify dementia-related proteins for replication analyses in the ARIC study. The first step was designed to filter out proteins with no or only weak association with cognitive decline rate after false discovery rate correction of 5%; we had a 80% statistical power to detect a beta of 0.08. The outcome, cognitive decline slope, is sensitive to any decrease in cognition, including both impairments that precede dementia and those not reaching levels that meet criteria for clinical dementia. The second step was therefore designed to identify from the proteins that passed the first step those that are also associated with dementia. We used a dementia outcome derived from electronic health records from the National Health Services (NHS) Hospital Episode Statistics (HES) database, an outcome with a high positive predictive value. The analysis model in the second step was the Cox proportional hazards model which had 80% power to identify a beta of 0.27 with a p-value<0.05. The replication analysis used the ARIC study, which used different dementia ascertainment methods, had a different dementia conversion rate, different participant characteristics at baseline, and originated from a different continent. Of the 21 proteins that passed first and second step in Whitehall II study, 15 (71%) were replicated in the ARIC study suggesting that our approach was successful within the limits of our dataset. Supplementary Figures 1a and b. Sample size needed to detect a range of beta coefficients using linear regression or Cox proportional hazards model. Graph in part a is derived using sample size 2242 (dashed line), power 0.8, p-value 0.002, and standard deviation of 1 for both the covariate of interest and for the error term. Graph in part b is derived using outcome number of 106 (dashed line), power 0.8, p-value 0.05, and standard deviation 1 for covariate of interest.

b) a)
Supplementary

Supplementary Figure 2. Associations between each of the 4953 proteins and rate of cognitive decline by -log10 p-values and beta coefficient.
Inverse rank-based normal transformed proteins plotted by their betas and p-values from linear regression in which rate of cognitive decline was the outcome. The 15 proteins associated with cognitive decline and dementia in Whitehall II study and with dementia in the ARIC study are labelled.

Supplementary Figure 3. Associations between each of the 4953 proteins and dementia by -log10 p-values and beta coefficient.
Inverse rank-based normal transformed proteins plotted by their betas and p-values from Cox proportional hazards model in which binary dementia was the outcome. The 15 proteins that associated with cognitive decline and dementia in Whitehall II study and with dementia in the ARIC study are labelled.

Supplementary Figure 4. Violin plots describing differences in standardized protein levels by dementia status.
All the 15 inverse rank-based normal transformed proteins were elevated in those who developed dementia. Blue, no dementia. Green, dementia.