Parkinson's disease determinants, prediction and gene-environment interactions in the UK Biobank

Abstract

Objective: To systematically investigate the association of environmental risk factors and prodromal features with incident Parkinson's disease (PD) diagnosis and the interaction of genetic risk with these factors. To evaluate whether existing risk prediction algorithms are improved by the inclusion of genetic risk scores.

Methods: We identified individuals with an incident diagnosis of PD (n=1276) and controls (n=500 406) in UK Biobank. We determined the association of risk factors with incident PD using adjusted logistic regression models. We constructed polygenic risk scores (PRSs) using external weights and selected the best PRS from a subset of the cohort (30%). The PRS was used in a separate testing set (70%) to examine gene-environment interactions and compare predictive models for PD.

Results: Strong evidence of association (false discovery rate <0.05) was found between PD and a positive family history of PD, a positive family history of dementia, non-smoking, low alcohol consumption, depression, daytime somnolence, epilepsy and earlier menarche. Individuals with the highest 10% of PRSs had increased risk of PD (OR 3.37, 95% CI 2.41 to 4.70) compared with the lowest risk decile. A higher PRS was associated with earlier age at PD diagnosis and inclusion of the PRS in the PREDICT-PD algorithm led to a modest improvement in model performance. We found evidence of an interaction between the PRS and diabetes.

Interpretation: Here, we used UK Biobank data to reproduce several well-known associations with PD, to demonstrate the validity of a PRS and to demonstrate a novel gene-environment interaction, whereby the effect of diabetes on PD risk appears to depend on background genetic risk for PD.

Figure 1

Associations of risk factors and incident cases of PD. Point estimates for association are depicted as log ORs and 95% CIs. Estimates of association were derived from logistic regression models adjusting for age, sex, Townsend deprivation index at recruitment and ethnicity. BMI, body mass index; PD, Parkinson’s disease.

Figure 2

PREDICT-PD determined probability (on the absolute risk scale) of PD, determined at recruitment, for individuals who would go on to develop PD (incident cases) and those who would not (controls). PD, Parkinson’s disease.

Figure 3

(A) Several candidate Polygenic Risk Scores (PRSs) were created using summary statistics from the Meta5 PD GWAS excluding UKB participants. For each candidate PRS, the degree of variation in PD risk explained was estimated using Nagelkerke’s pseudo-R2 metric. 95% CIs were derived from 1000 bootstrap resamples of the training dataset. As test statistics were approximately normally distributed, 95% CIs were derived from the normal distribution (mean±1.96 x SE). (B) Normalised PRS values for incident PD cases and controls. (C) OR of PD by PRS decile compared with lowest PRS decile. (D) Correlation between increasing PRS and earlier age at PD diagnosis. GWAS, genome-wide association studies; PD, Parkinson’s disease.

Figure 4

Interactions between risk factors for PD and the PD PRS were estimated using the attributable proportion due to interaction. Point estimates for the AP and 95% CIs are shown. PD, Parkinson’s disease; PRS, Polygenic Risk Score.