UK Biobank is a prospective study that enrolled more than 500,000 individuals aged 40 to 79 years from 22 evaluation centers across the United Kingdom between April 2006 to December 2010. During recruitment, all participants were assessed for demographic information, lifestyle factors, bodily measurements, and other health-related parameters by trained health professionals. Additionally, blood specimens were collected for genotyping. The UK Biobank study protocol is publicly available at https://www.ukbiobank.ac.uk/.

In this large population-based study of 502,461 participants, several exclusion criteria were applied to ensure data quality: (1) individuals with prevalent ADRD or PD at baseline; (2) those with missing data on urate levels, genetic information, and related covariates; (3) individuals with sex discordance; (4) outliers with genotype missingness or heterozygosity; (5) individuals with genetic kinship to other participants; and (6) individuals of non-European ancestry. As a result, a final sample of 382,182 participants was retained for the analysis. The flowchart is shown in Fig. S1.

The UK Biobank study was approved by the Northwest Multi-Center Research Ethics Committee, and each participant provided written informed consent before participating in the study. The data resource used for this study was obtained under application number 63,454 from the UK Biobank.

Baseline serum urate levels were measured using the uricase pedigree analysis package of the Beckman Coulter AU5800 platform (Randox Biosciences, Crumlin, UK). Participants were categorized into quartiles based on the distribution of urate levels according to sex. “Quartile 1” refers to the lowest 25% of participants with the lowest urate level, while “quartile 4” represents the highest 25% of participants with the highest urate level.

Neurodegenerative outcomes were identified using data on admissions and diagnoses with primary or secondary diagnosis based on the International Classification of Diseases (detailed information provided in Table S1) [19, 20]. The follow-up period ranged from March 16, 2006 to the end endpoint of follow-up (September 30, 2021 for centers in England; February 28, 2018, for centers in Wales; and July 31, 2021, for centers in Scotland). Person-years were calculated for each participant from the date of baseline assessment to the occurrence of neurodegenerative outcomes, death, or the end of follow-up, whichever occurred first.

Covariates possibly affecting the associations between urate levels and neurodegenerative outcomes, as indicated by previous studies, were taken into account in our analysis. A baseline touch-screen questionnaire was used to assess the potential confounding variables, including sociodemographic and lifestyle factors (e.g., age, sex, educational levels, smoking status, alcohol consumption and dietary habits), as well as personal and family history of diseases. Based on the baseline food frequency questionnaire, a diet score was calculated using the following elements: vegetables, fruits, fish, processed meat, unprocessed red meat, whole grains, and refined grains, as conducted in previous studies [21, 22]. Each diet factor received 1 point: consumption of at least 3 servings of vegetables per day, at least 3 servings of fruit per day, at least 2 servings of fish per week, no more than 1 serving of processed meat per week, no more than 1.5 servings of unprocessed red meat per week, at least 3 servings of whole grains per day, and no more than 1.5 servings of refined grains per week. The total diet score ranged from 0 to 7. Details of covariates were provided in Table S3.

The genotyping procedure and DNA array used in the UK Biobank study have been previously described [23]. In brief, each participant’s blood specimen was genotyped using the custom Affymetrix UK Biobank Axiom array. The genotyping data underwent phasing and imputation; SHAPEIT3 was used for phasing and IMPUTE3 was used for imputation, with a merged reference panel of UK10K and 1000 Genomes Phase 3 [24].

We used 20 independent single nucleotide polymorphisms (SNPs) (P < 5 × 10− 8, r2 < 0.1 within a 1000 kb window) identified in a genome-wide association analysis as genetic instruments in the MR (Table S2) [25]. These SNPs were used to construct the genetic risk score (GRS). The calculation of the GRS for each SNP involved coding them as 0, 1, or 2 based on the number of risk alleles, and each SNP was weighted by its relative effect size (β coefficient). The GRS for each individual was then obtained by summing the weighted scores using the PLINK “–score” command and the z-standardized value. The distribution of urate-related GRS is shown in Fig. S2. In this study, the genetic instrument showed a strong association with urate levels, with an F statistic of 173 and a P-value < 0.0001.

Baseline characteristics of the study population were outlined across quartiles of the urate levels, with continuous variables expressed as mean (standard deviation, SD) and categorical variables as percentages (%). Cox proportional hazard regression models were used to examine the associations of urate levels with neurodegenerative outcomes. Proportional hazards were tested using scaled Schoenfeld’s residuals. Three models were established: (1) model 1 adjusted for age, sex, and body mass index (BMI); (2) model 2 additionally adjusted for education levels, Townsend deprivation index, smoking status, and drinking status based on model 1; and (3) model 3 additionally adjusted for family history of diseases (hypertension, cardiovascular disease, and diabetes), healthy diet score, and personal history of diseases (kidney disease, hypertension, cardiovascular disease, and diabetes) based on model 2. The P-value for trend was calculated using the median value of urate in each quartile as a continuous variable [26]. Restricted cubic splines based on Cox proportional hazards regression model [27] were used to evaluate non-linear associations between urate levels and neurodegenerative outcomes in the multivariable model with 3 knots at the 25th, 50th, and 75th percentiles of the urate levels (with the minimum value used as the reference). To strengthen the robustness of the results, we performed several sensitivity analyses as follows: (1) excluded participants who had incident neurodegenerative outcomes at the initial 5 follow-up years to avoid reverse causality; (2) repeated the analysis after stratifying by age, sex, and BMI; (3) conducted Fine–Gray competing risk analysis to assess the competitive risk of non-neurodegenerative death [28]; and (4) divided the neurodegenerative death into deaths due to ADRD and PD respectively.

We employed both linear and non-linear MR methods to assess potential causal associations between urate levels and neurodegenerative outcomes. For the linear MR analyses, we examined the associations between urate-related GRS and neurodegenerative outcomes using a Cox regression model. The model was adjusted for various covariates, including age, sex, BMI, educational levels, Townsend deprivation index, smoking status, alcohol consumption, family history of diseases (hypertension, cardiovascular disease, and diabetes), healthy diet score, personal history of diseases (kidney disease, hypertension, cardiovascular disease, and diabetes), the first 10 principal components of ancestry, and genotype measurement batch. In the sensitivity analyses, (1) we employed an unweighted GRS model, calculated by summing the number of urate-related increasing alleles; (2) the SNP rs2231142, identified as the strongest in previous GWAS, was used as an instrumental variable to mitigate the potential introduction of horizontal pleiotropy [25]; and (3) the urate-related GRS was divided into quartiles to assess the linear MR results. In the non-linear MR analyses, we divided the sample into five strata based on the residual urate levels, which represented the differential value between the observed urate level and the genetically predicted urate level. Within each stratum, we evaluated the linear MR estimate, which contributed to the localized average causal effect (LACE) [29]. A meta-regression of LACE estimates against the mean of the exposure in each stratum was performed using a flexible semiparametric framework that applied the derivative of fractional polynomial models. This assessment aimed to determine whether a non-linear model offered a better fit for the LACE estimates compared to a linear model [30]. Two tests for non-linearity were conducted as follows: (1) a Cochran’s Q statistic to assess heterogeneity by analyzing differences between the LACE estimates, and (2) a trend test that involved meta-regression of LACE estimates against the mean value of urate in each stratum.

P-values were two-sided with < 0.05 defined as statistically significant. Statistical Analysis System 9.4 software for Windows was used to conduct the cohort analyses (SAS Institute Inc., Gary, NC, USA), and MR analyses were performed using R version 4.2.3 with “TwoSampleMR” and “NLMR” packages.