Plasma proteomics for risk prediction of Alzheimer's disease in the general population

Abstract

We aimed to develop and validate a protein risk score for predicting Alzheimer's disease (AD) and compare its performance with a validated clinical risk model (Cognitive Health and Dementia Risk Index for AD [CogDrisk-AD]) and apolipoprotein E (APOE) genotypes. The development cohort, consisting of 35,547 participants from England in the UK Biobank, was randomly divided into a 7:3 training-testing ratio. The validation cohort included 4667 participants from Scotland and Wales in the UK Biobank. In the training set, an AD protein risk score was constructed using 31 proteins out of 2911 proteins. In the testing set, the AD protein risk score had a C-index of 0.867 (95% CI, 0.828, 0.906) for AD prediction, followed by CogDrisk-AD risk factors (C-index, 0.856; 95% CI, 0.823, 0.889), and APOE genotypes (C-index, 0.705; 95% CI, 0.660, 0.750). Adding the AD protein risk score to CogDrisk-AD risk factors (C-index increase, 0.050; 95% CI, 0.008, 0.093) significantly improved the predictive performance for AD. However, adding CogDrisk-AD risk factors (C-index increase, 0.040; 95% CI, -0.007, 0.086) or APOE genotypes (C-index increase, 0.000; 95% CI, -0.054, 0.055) to the AD protein risk score did not significantly improve the predictive performance for AD. The top 10 proteins with the highest coefficients in the AD protein risk score contributed most of the predictive power for AD risk. These results were verified in the external validation cohort. EGFR, GFAP, and CHGA were identified as key proteins within the protein network. Our result suggests that the AD protein risk score demonstrated a good predictive performance for AD risk.

Keywords: Alzheimer's disease; dementia; enrichment analyses; protein–protein interaction network; proteomics.

FIGURE 1

The dose–response association between the Alzheimer's disease (AD) protein risk score and new‐onset AD in (a) the testing set of the development cohort and (b) the validation cohort. Adjusted for APOE genotypes and clinical risk factors in the CogDrisk‐AD including age, sex, education, obesity, diabetes, depression, high cholesterol, traumatic brain injury, smoking, loneliness, physical activity, cognitive activity, fish intake, hypertension, stroke, and occupational pesticides exposure.

FIGURE 2

The cumulative C‐index of the top 10 proteins with the largest absolute coefficients in the Alzheimer's disease (AD) protein risk score for new‐onset AD in the testing set of the development cohort and in the validation cohort. Based on the 31 candidate proteins, the top 10 proteins with the largest absolute coefficients were included in the AD prediction model sequentially (in order of coefficient from the largest to smallest).

FIGURE 3

The Gene Ontology (GO) function enrichment analyses, including (a) cellular component, (b) biological process and (c) molecular function, (d) Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses, and (e) protein–protein interaction (PPI) network of proteins derived from the AD protein risk score. The number beside each bar represents the number of observed proteins in each pathway. Statistical significance was defined as a two‐tailed p value<0.05 (dotted vertical line). Network nodes represent proteins and edges represent protein–protein associations. The larger‐sized node represents core proteins.