Genetic risk loci for Parkinson's disease and dementia in a large-scale population-based Taiwanese cohort (TPMI)

Summary

Background

The genetic architecture of Parkinson's disease (PD) and progression to PD dementia (PDD) remains incompletely characterized in Asians. Here, we investigated genetic risk factors for PD and PDD in Taiwanese individuals from the Taiwan Precision Medicine Initiative (TPMI), the largest non-European cohort integrating genetic and electronic medical record data.

Methods

We conducted a two-stage population-based case–control genome-wide association study (GWAS) with a 1:10 case-to-control ratio. Results were meta-analyzed with an independent Asian GWAS. We further constructed a polygenic risk score (PRS) to distinguish PD cases from controls. PDD risk was assessed using logistic regression and Cox models, with replication in an independent cohort that underwent whole-genome sequencing (WGS).

Findings

Among 463,447 TPMI participants, 4381 PD patients and 43,810 controls were analysed. We confirmed established PD loci at SNCA and LRRK2, and identified additional risk variants near HLA, AOAH–ELMO1, WFDC11–WFDC10B, TECPR1, BORCS7, and HIP1R. PRS models discriminated PD from controls with 70% accuracy. Among PD patients, 333 developed PDD. Genome-wide survival analysis identified PRDM15 rs141772267 as a novel PDD risk variant (HR = 4.20; 95% CI 2.67–6.60; P = 5.41 × 10⁻¹⁰), together with two loci near CNOT6L–MRPL1 and NPY2R–MAP9. Carriers of two or more minor risk alleles showed a markedly increased risk of progression to PDD (P = 2 × 10⁻¹⁶), which was replicated in an independent WGS cohort.

Interpretation

Our findings highlight both shared and ethnicity-specific PD risk loci and identify PRDM15 as a potential novel contributor to PDD. Further multi-ethnic and functional studies are warranted.

Funding

Academia Sinica and National Development Fund.

Research in context.

Evidence before this study

We conducted a literature search to identify genome-wide association studies evaluating the risk variants contributing to Parkinson's disease and its progression to Parkinson's disease dementia in Asians. We searched PubMed using the terms “genome-wide association studies” “Parkinson's disease” “Parkinson's disease dementia” “Asians” covering the period from database inception to 1st June 2025, with no language restrictions. We identified six genome-wide association study (GWAS) investigating genetic risk factors for Parkinson's disease (PD) in Asian populations. Only two pilot studies included data on progression to Parkinson's disease dementia (PDD), each with limited sample sizes and without replication cohorts.

Added value of this study

In this GWAS involving 463,447 participants and a 1:10 case-to-control ratio, followed by replication and combined analyses, we characterized both shared and population-specific genetic architecture of PD in a large Taiwanese cohort, an underrepresented East Asian population. Leveraging the longitudinal follow-up data from the Taiwan Precision Medicine Initiative (TPMI), the largest non-European cohort integrating genotypic data with electronic medical records, we identified a novel risk locus for PDD in PRDM15 (rs141772267), which conferred more than a fourfold increased risk of dementia progression among PD patients. We also identified two intergenic variants (CNOT6L–MRPL1 and NPY2R–MAP9) associated with PDD. Carriers of two or more minor alleles across these loci exhibited a synergistically elevated risk of developing PDD, a finding replicated in an independent cohort.

Implications of all the available evidence

Our results highlight both shared and ethnicity-specific PD risk loci through Asian meta-analysis and identify PRDM15 as a potential novel genetic contributor to PDD risk, along with two additional variants near CNOT6L–MRPL1 and NPY2R–MAP9. In a largely uncharacterized East Asian population, individuals carrying two or more minor risk alleles had a markedly higher likelihood of progression to PDD. These findings underscore the importance of including diverse populations in genetic studies and warrant further multi-ethnic and functional investigations to elucidate the biological mechanisms underlying these associations.

Introduction

Parkinson's disease (PD) is a common neurodegenerative disorder characterized by dopaminergic neuron loss in the substantia nigra, and Lewy body formation. As the pathology advances from the brainstem to the cortex, patients exhibit various motor and non-motor features, with one of the most challenging being the development of Parkinson's disease dementia (PDD).¹ Investigations of the genetic architecture of PD and its underlying biological pathways have highlighted its heterogeneity. Many genome-wide association study (GWAS) have explored genetic variants underlying cognitive impairment in PD; however, only a limited number of risk variants have been identified, and most studies have been conducted in cohorts of European ancestry.2, 3, 4, 5 There remains a need to identify risk factors for PD development and progression in more diverse populations, to inform the design of future clinical trials aimed at slowing disease progression.

Genetic studies of early-onset and familial PD have advanced molecular insights, while GWASs have revealed common variants associated with idiopathic PD, the most prevalent form of PD.6, 7, 8, 9, 10, 11 A meta-analysis of 17 GWAS datasets from individuals of European ancestry identified 90 genetic risk variants, explaining 16–22% of PD heritability.⁷ Studies in non-European populations remain limited.^6,8,9,11 Notably, genetic risk factors for PD differ between Asian and non-Asian populations. For example, Asian populations rarely exhibit the MAPT H2 haplotype and LRRK2 p.G2019S variant,^6,8,9 or the GBA1 rs3115534 variant, which is associated with PD risk and age at onset in African populations.⁹ On the other hand, the LRRK2 p.R1628P and p.G2385R variants are predominantly observed in Asian populations,⁶ and GWASs in Asian populations have identified novel risk loci not observed in European cohorts,^6,8, 9, 10, 11 including SV2C, WBSCR17, SMPD1 and HEATR6.^8,11,12 A recent multi-ancestry meta-analysis reported 78 significant PD loci, including 12 potentially novel genes, but the majority of participants were of European ancestry.¹⁰ Large-scale initiatives are underway to sequence genomic data from underrepresented populations, with the aims of identifying novel loci, refining known associations, and reducing disparities in precision medicine.¹³ As PD therapeutics move toward targeted clinical trials, it is critical to elucidate the genetic architecture of PD in underrepresented groups.

PD studies among Taiwanese participants are relatively scarce.6, 7, 8, 9, 10, 11 Taiwan's diverse population comprises Han Chinese and indigenous Malayo-Polynesian communities, as well as Southeast Asian and Austronesian immigrants. While Taiwan is home to a culturally diverse population, most participants in our study are of Han Chinese ancestry. The Taiwan Precision Medicine Initiative (TPMI) was designed to establish a large-scale cohort of individuals of Han Chinese ancestry, among whom to perform genetic profiling using population-optimized SNP arrays. As the largest non-European precision medicine cohort to date, the TPMI integrates comprehensive genetic data with electronic medical records (EMRs) from 463,447 participants,^14,15 providing a unique opportunity to investigate PD risk in Taiwan. Here we report our use of this population-based resource to perform a GWAS aiming to characterize genetic factors contributing to PD and its progression to PDD.

Methods

Study participants and phenotyping

We analyzed the TPMI dataset, which integrates extensive EMR data with genetic information from 463,447 individuals. The TPMI dataset has previously been described in detail.^14,15 Briefly, the TPMI is a large population-based cohort established by Academia Sinica and 16 major medical centres in Taiwan, which includes >560,000 participants with genome-wide genotyping and longitudinal EMRs.^14,15 The dataset shows high completeness (>250,000 participants with ≥5 years of follow-up, and >70,000 with ≥10 years of follow-up), broad genomic coverage (>98% for variants common among Han Chinese), and robust validation metrics (mean imputation quality score [INFO] = 0.91; >96% of variants with INFO >0.7; and >99.8% concordance with WGS data). The TPMI captures all major geographic regions and Han Chinese subgroups, providing comprehensive representation of a genetically homogeneous portion of the Taiwanese population.

To identify genetic variants associated with PD, a two-stage case–control GWAS was conducted, including a discovery cohort and a replication cohort (Fig. 1A). PD was defined based on the International Classification of Disease, Tenth Revision, Clinical Modification (ICD-10-CM) code G20.0, as diagnosed and coded by board-certified neurologists. All data were linked to the EMR of participating medical centers in the TPMI.^14,15 The EMR contained detailed clinical information, including demographic data, clinical visits, medication responses, neuroimaging reports, hospitalization records, and physician-confirmed disease diagnoses coded following ICD standards. To minimize diagnostic misclassification, the present analysis excluded individuals with potentially atypical parkinsonian syndromes, such as multiple system atrophy, corticobasal syndrome, progressive supranuclear palsy, and vascular or drug-induced parkinsonism, based on clinical documentation and diagnostic codes. Prior evaluation of the validity of using ICD codes for PD identification in Taiwanese tertiary medical centers has shown 94% specificity when combined with neurologist-reviewed medical records.¹⁶

Fig. 1.

Flowchart of participant enrollment and Manhattan plot for the associations of SNPs with Parkinson's disease (PD) in the TPMI cohort. (A) Screening process for people with PD, progression to Parkinson's disease dementia (PDD), and age- and sex-matched controls in the discovery and replication cohorts of TPMI dataset. An independent whole-genome sequencing-based replication cohort was used to validate the associations of genetic variants with PDD. GWAS, genome-wide association study; GWSS, genome-wide survival study. (B) Manhattan plot for the associations of SNPs with PD in the combined TPMI database. The horizontal axis shows the chromosomal position, and the vertical axis shows the significance of tested markers from 6,682,780 polymorphic SNPs in 4381 PD cases and 43,810 controls. The threshold for genome-wide significance (P < 5 × 10⁻⁸) is indicated by a gray dashed line. (C) Quantile–quantile plot of results from logistic regression with sex, age, medical comorbidities of diabetes mellitus, hypertension, and hyperlipidemia, PC1-PC10. (D and E) Regional association plots of genetic loci associated with PD identified by TPMI genome-wide association study, encompassing SNCA (D) and LRRK2 (E). The x-axis shows chromosomal positions, the left y-axis shows –log₁₀P values from the GWAS, and the gray dashed line indicates the significance threshold P = 5 × 10⁻⁸.

The patient inclusion criteria were age of >40 years, and a ICD-10-CM diagnosis code of G20.0. The patient exclusion criteria were atypical or secondary parkinsonism, eg, multiple system atrophy, progressive supranuclear palsy, and vascular or drug-induced parkinsonism (ICD-10-CM codes G21, F20.9, and I67.9.0, respectively); history of neuroleptic exposure, or cerebrovascular events preceding parkinsonian features. The control group included individuals aged >40 years, with no ICD-10-CM codes for PD or other neurodegenerative disorders. Controls were excluded if they had ever been prescribed anti-parkinsonian medications (amantadine, bromocriptine, carbidopa, levodopa, pergolide, pramipexole, rasagiline, ropinirole, rotigotine, or selegiline), or if they were related to other study participants within three generations (first-, second- or third-degree relatives). After stringent sample-level quality control, the analysis included 4381 individuals with PD and 273,462 controls.

The discovery cohort included participants recruited from hospitals in northern Taiwan, whereas the replication cohort comprised participants from southern Taiwan. This stratification by geographic region provided an internal validation framework within the same ethnic background, thereby minimizing the risk of false-positive associations arising from population substructure or regional recruitment bias. To enhance the robustness and reproducibility of the findings, all association results were also evaluated in the combined analysis of discovery and replication cohorts.

The participants with PD who later developed dementia, defined by a Mini-Mental State Examination (MMSE) score of <26, after excluding secondary causes of cognitive impairment, were classified as having Parkinson's disease dementia (PDD). This MMSE cut-off was based on the criteria established by Taiwan's National Health Insurance (NHI) program, which governs the reimbursement of acetylcholinesterase inhibitors for patients with PDD. The NHI program is a mandatory social insurance scheme covering approximately 99.99% of Taiwan's population, thereby ensuring comprehensive and consistent diagnostic and treatment documentation across all participating medical centres. To ensure diagnostic specificity, our analyses excluded individuals with other types of dementia, including Alzheimer's disease (AD), frontotemporal dementia, hypothyroidism, neurosyphilis, brain tumours, and systemic diseases causing cognitive decline. As comparison controls for secondary analyses, age- and sex-matched PD patients without dementia were selected at a 1:6 ratio.

An independent cohort of PD patients, who were not enrolled in the TPMI, was recruited from the Movement Disorder Clinics of National Taiwan University Hospital between 2015 and 2024. These individuals underwent whole-genome sequencing (WGS) and served as an ethnicity-matched replication cohort (Fig. 1A).^13,17 Participants in the WGS validation cohort were investigated for genetic relatedness with those enrolled in the TPMI cohort. Genome-wide relatedness analysis was performed to identify individuals showing kinship within three generations, who were excluded from the WGS validation dataset to ensure sample independence.

Genotyping and imputation

The detailed genotyping, imputation, and quality control procedures have previously been described.^14,15 Briefly, to ensure optimal coverage of the Han Chinese population, genotyping was performed using two customized high-density Axiom single-nucleotide polymorphism (SNP) arrays (TPMv1 and TPMv2; Thermo Fisher, Waltham, MA, USA). The first-generation TPMv1 array contained 686,463 SNPs, and was designed to maximize imputation accuracy for both common and low-frequency variants, based on whole-genome sequencing data from the Taiwan Biobank.¹⁸ The updated TPMv2 array comprised 743,227 SNPs, and was developed based on insights from the initial phase of the project, to further improve genomic coverage and inclusion of population-specific and disease-associated variants. Each array was curated to incorporate previously described risk loci and variants relevant to complex diseases in East Asian populations, enabling comprehensive genetic characterization of the Taiwanese population. Following quality control, 401,710 genetic variants and 463,447 Taiwanese participants were retained for downstream analyses. Genome imputation was conducted in IMPUTE5, using a reference panel of 1498 whole-genome sequences from the Taiwan Biobank.¹⁸ Additionally, a chip-GWAS was performed to minimize the bias from different chips, yielding a dataset of 8,046,864 well-imputed common genetic variants.

Quality control

The genetic ancestry inference procedure has previously been described in detail.¹⁴ Briefly, we selected SNPs using the following criteria: call rate >0.95, minor allele frequency >0.01, Hardy–Weinberg equilibrium P > 1 × 10⁻⁴, and exclusion of INDELs. Subjects were removed if fewer than 95% of genotypes were called. Genetically inferred ancestry was determined using principal component projection analysis, and a supervised ADMIXTURE model,¹⁹ with the 1000 Genomes Project serving as the reference panel.²⁰ People who were genetically similar to the Han Chinese reference population were retained for further analysis. KING software was used to identify and exclude duplicates or close relatives, as determined by cryptic relatedness (PropIBD cut-off: 0.1875). Genetic relatedness analysis was performed, and participants showing kinship within three generations (first-, second-, or third-degree relatives) were excluded. When multiple affected family members were identified, only one index case was retained, and their relatives were excluded from the control group to avoid genetic overlap. Next, imputation was performed, and variants were further filtered in PLINK if they met any of the following conditions: minor allele frequency ≤0.001, imputation INFO score ≤0.8, call rate ≤0.95, or Hardy–Weinberg equilibrium P ≤ 1 × 10⁻⁴.

After applying these stringent quality control steps,^14,15 the combined dataset included 6,682,780 variants, achieving an average call rate of 98.91% ± 1.21%. With adjustment for age, sex, comorbidities (ie, diabetes, hypertension, and hyperlipidaemia), and the first 10 principal components (PC1–PC10), the quantile–quantile plot of logistic regression P values yielded a genomic inflation factor (λ) of 1.033 (Fig. 1C). Linkage disequilibrium (LD) score regression revealed genetic covariance intercepts that were close to zero or within the 95% confidence interval (CI), indicating no overlap between cohort samples.

Genome-wide association study for PD

A two-stage case–control GWAS was designed to identify genetic variants associated with PD, utilizing a discovery cohort, a replication cohort, and a subsequent combined analysis. To leverage a modest increase in statistical power, a 1:10 case-to-control ratio was applied. SNPs that achieved replication significance in the discovery stage (P < 5 × 10⁻⁵), and were successfully replicated in the validation cohort, were considered PD risk variants. To account for potential confounding and population stratification effects, all association analyses were adjusted for age, sex, common comorbidities, and the first ten principal components (PC1–PC10).

Conditional analysis

Conditional analyses were performed for each previously defined locus using the PLINK2 “--condition” flag. Association tests were based on a generalized linear model, adjusting for all covariates mentioned in the PD GWAS and the genotype of the lead SNP. Linkage disequilibrium (LD) was assessed by calculating the pairwise squared correlation (r²) within a window of 4000 SNPs using 1496 whole-genome sequences from Taiwan Biobank. An SNP was considered independent of the lead SNP if r² ≤ 0.1.

Genome-wide association study for PDD and survival analysis

To elucidate genetic factors associated with cognitive decline in PD, we further conducted a GWAS of PDD. To identify genetic variants associated with PDD among individuals with PD, we applied a two-stage case–control GWAS design, including a discovery cohort, a replication cohort, and a subsequent combined analysis (Fig. 1A). A 1:6 ratio of cases to age- and sex-matched controls was used for this analysis and for logistic regression, owing to the limited number of patients who progressed to PDD and the availability of matched PD patients without dementia during follow-up. Analyses were performed in PLINK (v.1.9), with adjustment for age at onset, disease duration, type 2 diabetes, hypertension, stroke, and PC1–PC5. No variants reached genome-wide significance (P < 5 × 10⁻⁸); however, several showed suggestive associations (P < 5 × 10⁻⁶).

Progression from PD with normal cognitive function to PDD was analyzed using time-to-event methods. In the Cox proportional hazards models, the time-to-event variable was defined using the TPMI framework.^14,15 The time origin was defined as the date of PD diagnosis (left-censored), and the event was the first clinical diagnosis of PDD (right-censored). Patients who did not develop dementia during the observation period were right censored at their last follow-up or death, whichever occurred first.

Kaplan–Meier survival curves were generated to estimate the cumulative probability of dementia-free survival among patients with PD, stratified by genotype groups (eg, risk allele carriers vs. non-carriers). The uncertainty around the survival estimates was illustrated by plotting 95% confidence intervals (CIs). Between-group differences in survival distributions were compared using the log-rank test.

Meta-analysis

To investigate variants with potential linkage to PD in Asian populations, we performed a meta-analysis of our dataset and a previously published Asian GWAS.⁸ Both the combined analysis of the discovery and replication cohorts, and the meta-analysis with the external GWAS were conducted using an inverse-variance–weighted fixed-effects model implemented in PLINK.

Polygenic risk score model

Based on the summary statistics, we developed a polygenic risk score (PRS) for PD, using the PRS-CS method, which implements a Bayesian regression framework with continuous shrinkage priority to estimate posterior SNP effect sizes. PRS-CS assumes that the true effects of genetic variants follow a mixture distribution—with most variants having effect sizes near zero, while a small subset has larger nonzero effects. Linkage disequilibrium information was derived from East Asian populations in the 1000 Genomes project.²⁰ We used the auto mode, in which the global shrinkage parameter is adaptively learned from the data. To evaluate the model, we used an independent validation set comprising 603 PD cases and 6030 matched controls. Model fitting was determined using an area under the receiver operating characteristic (ROC) curve (AUC). In addition to the PRS, we incorporated clinical demographic variables (age and sex) and metabolic comorbidities (diabetes mellitus, hypertension, and hyperlipidemia), which increase the risk of cognitive decline and dementia,²¹ along with the first ten principal components (PC1–PC10).

Gene expression and interactome analysis

We analyzed gene expression levels across 54 specific tissue types from the Genotype-Tissue Expression project, using FUMA (v.1.6.2). Protein–protein interactions were extracted from the Protein Interaction Network Analysis platform (v.3.0), and visualized as a network using Cytoscape (v.3.10.2). For the identified proteins, functional annotation was performed using the Molecular Signatures Database (MSigDB, v.7.4).

Power analysis

Power calculations for the genetic association study were performed using the Genetic Association Study (GAS) Power Calculator. This analysis was designed to estimate the statistical power to detect associations between genetic variants and the trait of interest under a GWAS framework. Input parameters included minor allele frequency (MAF), effect size, sample size, and significance threshold (α = 5 × 10⁻⁸ for genome-wide significance). The calculations assumed an additive genetic model and were used to determine the minimum detectable effect size with 80% power, as well as to evaluate whether the available sample size was adequate to detect common variants with moderate effects.

Statistical analysis

Demographic and clinical characteristics were summarized as mean ± standard deviation (SD) for continuous variables and as counts (percentages) for categorical variables. Between-group comparisons were performed using Student's T-test for continuous variables and the chi-square test for categorical variables. All statistical analyses were performed using R software (version 4.3.2; R Foundation for Statistical Computing, Vienna, Austria). For comparisons of demographic and clinical characteristics, a two-sided P value of <0.05 was considered statistically significant.

Ethics approval

This study was approved by the Institutional Review Boards of National Taiwan University Hospital (201912110RINC, 202405106RINB, and 202409029RINC) and Academia Sinica, Taiwan (AS-IRB01-18079). All collected data were de-identified before statistical analysis. All ethical regulations relevant to human research participants were followed.

Role of funders

The funding sources were not involved in the study design, data collection, data analysis, data interpretation, or manuscript preparation.

Results

Baseline characteristics of the study participants

Among the 463,447 eligible participants in the TPMI, we identified 4381 individuals who were >40 years old and had PD (ICD-10-CM code G20.0) (Fig. 1A). For the control group, we selected 273,462 individuals who were >40 years old; did not have the ICD-10-CM codes G20.0, G21, F20.9, or I67.9.0; and had not received treatment with anti-parkinsonism medications. For the two-stage case–control GWAS, to investigate PD-linked genetic variants, a discovery cohort was formed with participants from hospitals in northern Taiwan (2240 PD cases and 22,400 age- and sex-matched controls), and a replication cohort with participants from hospitals in southern Taiwan (2141 PD cases and 21,410 age- and sex-matched controls). The combined analysis included a total of 4381 individuals with PD and 43,810 controls (Fig. 1A). Compared to individuals with PD, those without PD exhibited higher rates of diabetes and hypertension, and lower rates of hyperlipidemia (Table 1).

Table 1.

Demographics of participants in combined analysis, discovery and replication cohorts.

	People with PD	People without PD	P value
Combined analysis
Number	4381	43,810
Age, y, mean ± SD (range)	73.36 ± 10.19 (40–103)	73.36 ± 10.15 (40–104)	0.9938
Sex, male, n (%)	2307 (52.66%)	23,110 (52.75%)	0.9081
DM, n (%)	1293 (29.5%)	15,290 (34.9%)	<0.0001
Hypertension, n (%)	1990 (45.42%)	21,155 (48.29%)	0.0003
Hyperlipidemia, n (%)	1000 (22.83%)	8085 (18.45%)	<0.0001
Discovery cohort
Number	2240	22,400
Age, y, mean ± SD (range)	73.49 ± 10.23 (40–103)	73.49 ± 10.22 (40–103)	0.9967
Sex, male, n (%)	1186 (52.95%)	11,863 (52.96%)	0.9902
DM, n (%)	606 (27.05%)	7875 (35.16%)	<0.0001
Hypertension, n (%)	992 (44.29%)	10,801 (48.22%)	0.0004
Hyperlipidemia, n (%)	507 (22.63%)	3859 (17.23%)	<0.0001
Replication cohort
Number	2141	21,410
Age, y, mean ± SD (range)	73.23 ± 10.15 (40–102)	73.22 ± 10.14 (40−104)	0.9945
Sex, male, n (%)	1121 (52.36%)	11,247 (52.53%)	0.8787
DM, n (%)	687 (32.09%)	7415 (34.63%)	0.0181
Hypertension, n (%)	998 (46.61%)	10,354 (48.36%)	0.1230
Hyperlipidemia, n (%)	493 (23.03%)	4226 (19.74%)	0.0003

SD, standard deviation; DM, diabetes mellitus; PD, Parkinson's disease.

GWAS for PD risk

Two genomic loci, SNCA and LRRK2, harbored multiple SNPs showing robust evidence of replication, reaching P < 5 × 10⁻⁵ in both the discovery and replication cohorts. In the combined analysis, these variants achieved genome-wide significance (P < 5 × 10⁻⁸) after adjustment for age, sex, common comorbidities, and PC1–PC10 (Fig. 1B). The quantile–quantile plot of logistic regression P values yielded a genomic inflation factor (λ) of 1.033 and a pseudo-R² value of 0.006 (Fig. 1C). Combined analysis revealed the following key variants: rs2619341 (SNCA, c.306 + 1624 T > G; odds ratio [OR] 1.29, 95% CI 1.14–1.30, P = 9.04 × 10⁻⁹), rs146373824 (LRRK2, c.6109 + 1035 T > C; OR 1.59, 95% CI 1.41–1.79, P = 8.38 × 10⁻¹⁴), and rs34778348 (LRRK2, c.7153G > A, p.G2385R; OR 1.52, 95% CI 1.35–1.27, P = 4.93 × 10⁻¹²; Fig. 1D and E and Table 2). The three variants listed in the combined analysis were selected based on the following criteria: (1) they reached genome-wide significance (P < 5 × 10⁻⁸) in the combined analysis; (2) they showed consistent direction and magnitude of effect in both the discovery and replication cohorts; and (3) they correspond to the lead (index) SNPs after linkage disequilibrium–based clumping. Specifically, rs2619341 represents the lead SNP at the SNCA locus, while rs146373824 and rs34778348 (p.G2385R) are correlated variants capturing the same association signal at the LRRK2 locus. These variants are reported to transparently summarize the strongest genome-wide significant signals identified in the combined analysis.

Table 2.

Lead SNPs of genome-wide significant loci associated with PD in the combined analysis.

Chr	SNP	Position	Gene	Minor allele	Stage	MAF (case, control)	TWB (n = 1492)	OR	95% CI	P valuea
4	rs2619341	89,820,622	SNCA	G	Discovery	(0.87, 0.84)	0.1588	1.22	1.11–1.33	2.17 × 10−5
					Replication	(0.87, 0.85)		1.20	1.10–1.33	5.60 × 10−5
					Combined	(0.87, 0.85)		1.20	1.14–1.30	9.04 × 10−9
12	rs146373824	40,341,489	LRRK2	T	Discovery	(0.04, 0.02)	0.0265	1.62	1.37–1.91	1.30 × 10−8
					Replication	(0.03, 0.02)		1.56	1.31–1.87	1.09 × 10−6
					Combined	(0.04, 0.02)		1.59	1.41–1.79	8.38 × 10−14
12	rs34778348	40,363,526	LRRK2	G	Discovery	(0.04, 0.02)	0.0268	1.54	1.31–1.82	2.04 × 10−7
					Replication	(0.04, 0.02)		1.50	1.26–1.79	5.00 × 10−6
					Combined	(0.04, 0.02)		1.52	1.35–1.72	4.93 × 10−12

SNP, single nucleotide polymorphism; Chr, chromosome; MAF, minor allele frequency; TWB, Taiwan Biobank; OR, odds ratio; CI, confidence interval.

^aThe P values of logistic regression adjusting for age, co-morbidities of DM, hypertension, hyperlipidemia and the PC1–PC10.

A subsequent LD analysis was performed to evaluate the relationship between the two identified LRRK2 variants, rs146373824 (c.6109 + 1035 T > C) and rs34778348 (c.7153G > A, p.G2385R). LD analysis based on the TPMI cohort and the Taiwan Biobank reference panel showed strong linkage (r² > 0.8). Conditional analysis using rs146373824 as a covariate confirmed that rs34778348 was not independently associated with PD risk (P = 0.5884), indicating that both variants represent the same association signal within the LRRK2 locus.

Meta-analysis of GWASs

The TPMI-derived dataset did not overlap with the Taiwanese participants from previous Asian GWASs (Supplementary Table S1).⁸ To discover novel potential PD loci specific to Asian populations, we performed a meta-analysis in PLINK, integrating our results with those of a previous Asian GWAS.⁸ This combined analysis included 79,766 individuals (11,105 PD cases and 68,661 controls) and revealed several loci exceeding the genome-wide significance threshold (P < 5 × 10⁻⁸; Fig. 2A), most of which had been previously reported.2, 3, 4, 5, 6, 7, 8, 9, 10, 11 The most prominent signals, spanning ASH1L, PARK16, ITPKB, MCCC1, FM47E, SCARB2, SNCA, SV2C, and LRRK2, remained consistent and even more robust in the merged meta-analysis (Table 3).

Fig. 2.

Manhattan plot for the meta-analysis of the genome-wide association study (GWAS) of Parkinson's disease (PD) in the merged Asian dataset. (A) The merged Asian sample combined the TPMI dataset and an earlier Asian GWAS showing SNPs associated with PD. The threshold for genome-wide significance (P < 5 × 10⁻⁸) is indicated by a gray dashed line. (B–G) Regional association plots for the top significant loci in the merged Asian meta-analysis, including intergenic variants between HLA-DRB1 and HLA-DQA1 (B) and between AOAH and ELMO1 (C), intronic variants in TECPR1 (D), BORCS7 (E), and HIP1R (F), and an intergenic variant between WFDC11 and WFDC10B (G).

Table 3.

Association of novel and previously reported PD risk loci in the merged Asian GWAS meta-analysis.

Chr	Position	SNP	Gene	Minor allele	MAF (cases; controls)	Functional consequence	P1a	P2b	P3c	P4d	Direction	I2, %	P for heterogeneity
1	155,376,394	rs138045074	ASH1L	C	(0.04, 0.03)	intronic	5.86E-05	6.78E-05	2.33E-08	2.33E-08	–	0	0.3769
1	205,784,749	rs5014256	RAB29; PARK16	C	(0.44, 0.46)	intergenic	0.00576	4.59E-19	2.74E-16	0.06417	++	95.1	6.84E-06
1	226,659,011	rs16846351	ITPKB	G	(0.07, 0.06)	ncRNA_intronic	0.01053	8.16E-10	8.63E-10	0.01702	–	85.1	0.009669
3	183,012,503	rs9858038	MCCC1	G	(0.42, 0.42)	intergenic	0.0814	1.25E-16	1.92E-12	0.1295	–	95.4	2.81E-06
4	947,502	rs6815138	TMEM175	C	(0.35, 0.37)	intronic	0.0002168	0.0002807	2.16E-07	2.16E-07	++	0	0.9774
4	15,751,676	rs10000290	BST1;CD38	G	(0.33, 0.34)	intergenic	0.009968	5.29E-05	2.91E-06	8.50E-06	++	9.4	0.2935
4	76,191,731	rs28465660	FM47E-SCARB2	T	(0.35, 0.37)	intronic	0.002551	4.49E-08	2.10E-09	0.000728	–	68.4	0.07543
4	89,762,971	rs34806123	SNCA	A	(0.44, 0.47)	intronic	1.41E-07	2.04E-37	4.18E-37	0.01777	–	96.6	6.83E-08
5	76,274,833	rs1995381	SV2C	G	(0.11, 0.09)	intronic	2.26E-05	4.32E-07	5.47E-11	5.47E-11	–	0	0.4939
6	27,275,701	rs6917419	PRSS16	T	(0.15, 0.14)	intergenic	0.008488	5.84E-06	4.75E-07	0.000433	++	51.6	0.1505
6	32,605,785	rs602457	HLA-DRB1; HLA-DQA1	C	(0.07, 0.08)	intergenic	0.00267	7.03E-05	9.67E-07	9.67E-07	++	0	0.3636
6	111,849,088	rs12661662	FYN	A	(0.16, 0.17)	intronic	0.3104	4.78E-08	5.67E-06	0.1497	–	90.2	0.001375
7	36,845,635	rs34182751	AOAH;ELMO1	C	(0.03, 0.02)	intergenic	2.75E-06	0.008433	1.38E-07	9.41E-07	–	12.3	0.2856
7	71,296,424	rs28479972	WBSCR17	A	(0.48, 0.47)	intronic	0.05855	3.79E-08	1.66E-07	0.0394	++	84.3	0.01161
7	98,239,415	rs6963518	TECPR1	C	(0.46, 0.49)	intronic	2.63E-05	0.008149	1.28E-06	1.09E-05	++	17.5	0.2708
10	102,875,591	rs12765002	BORCS7	T	(0.43, 0.45)	ncRNA_intronic	0.0001756	0.001915	1.24E-06	1.24E-06	–	0	0.6687
11	83,770,514	rs75208430	DLG2	C	(0.05, 0.05)	intronic	0.05767	1.16E-05	8.38E-06	0.01019	–	66.5	0.08404
12	40,341,489	rs146373824	LRRK2	C	(0.04, 0.02)	intronic	8.38E-14	1.21E-23	1.21E-34	9.34E-08	–	81.4	0.02029
12	122,842,051	rs10847864	HIP1R	G	(0.44, 0.46)	intronic	3.72E-05	0.00142	2.11E-07	2.11E-07	++	0	0.6077
15	61,712,149	rs2261069	VPS13C	T	(0.06, 0.07)	intergenic	0.001828	0.0001267	8.50E-07	8.50E-07	–	0	0.6865
18	43,098,270	rs4130047	RIT2	C	(0.39, 0.37)	intronic	0.02063	5.04E-08	4.17E-08	0.01387	–	79.9	0.02561
20	45,671,332	rs115511611	WFDC11; WFDC10B	A	(0.02, 0.03)	intergenic	0.0002606	0.001755	1.75E-06	1.75E-06	–	0	0.6049

^aP1 represents the P values of logistic regression with age, sex, DM, HTN, HLP, PC1-PC10 of the current TPMI study.

^bP2 represents the P values of meta-analysis of a previous Asian GWAS (Foo et al., 2020).

^cP3 represents the P values of meta-analysis of the previous Asian GWAS (Foo et al., 2020) and the current TPMI data using PLINK.

^dP4 represents the P values of random-effects meta-analysis of the previous Asian GWAS (Foo et al., 2020) and the current TPMI data using PLINK.

This combined Asian meta-analysis integrated results across independent cohorts with differing study designs and sample ascertainment. In this context, in addition to genome-wide significant loci (P < 5 × 10⁻⁸), we also examined variants reaching P < 5 × 10⁻⁶ as suggestive associations, with the aim of highlighting potentially informative signals that showed consistent effects across datasets but did not uniformly achieve genome-wide significance. Accordingly, additional robust signals (P < 5 × 10⁻⁶) included previously reported loci (TMEM175, BST1, PRSS16, WBSCR17, VPS13C, and RIT2) and novel loci, including intergenic variants between HLA-DRB1 and HLA-DQA1, between AOAH and ELMO1, and between WFDC11 and WFDC10B; and intronic variants in TECPR1, BORCS7, and HIP1R (Fig. 2A and Table 3). Notably, rs10847864 in HIP1R was recently identified as a novel PD risk variant in a multi-ancestry GWAS.¹⁰ Regional association plots of potentially novel loci are shown in Fig. 2B–G and those of previously reported loci are illustrated in the Supplementary Fig. S1A–H. In addition to the fixed-effects model, we performed a random-effects inverse-variance meta-analysis (DerSimonian–Laird). Most loci exhibited low heterogeneity (I² < 30%, P > 0.1), with highly concordant results between the two models (Table 3). Among the few loci showing moderate-to-high heterogeneity (I² ≥ 50%), random-effects P values were slightly attenuated but remained directionally consistent. Importantly, all key loci, including SNCA, LRRK2, TMEM175, and PARK16, remained significant under both models.

In our present study along with GWASs conducted in European, Ashkenazi Jewish, and African populations,^7,9,10,22 a total of 262 genome-wide significant variants have been summarized (Supplementary Table S2). Among these, 145 variants overlapped, and 84 showed concordant directions of effect. Several independent risk genes previously reported in other populations (eg, SNCA, LRRK2, and TMEM175) were identified as independent signals in our merged Asian cohort, albeit represented by different lead variants. These similarities underscore shared pathogenic pathways underlying PD across diverse ancestries. In contrast, the GBA1 rs3115534 variant, which confers PD risk and earlier age of onset in African populations,⁹ was not detected in our cohort. On the other hand, several prominent signals detected in our merged Asian dataset have not been observed in European, Ashkenazi Jewish, or African studies, including variants in WBSCR17, intergenic variants between HLA-DRB1 and HLA-DQA1, between AOAH and ELMO1, and between WFDC11 and WFDC10B; and intronic variants in TECPR1, BORCS7, and HIP1R. Together, these findings highlight both shared biological mechanisms and ancestry-specific genetic architectures contributing to PD risk across global populations.

Association between PRS and risk of PD

To further evaluate the effects of the identified Asian-specific variants contributing to PD, we calculated three sets of PRS. By combining the PRS with the clinical characteristics, encompassing age, sex, medical co-morbidities and the top 10 principal components, the AUC for discriminating PD patients from controls was 0.7051 (95% CI 0.6734–0.7369) using the TPMI-based model and 0.71307 (95% CI 0.6816–0.7446) using the merged Asian meta-analysis result (Supplementary Fig. S2). When applying the model based on clinical characteristics alone, the AUC was 0.7008 (95% CI 0.6689–0.7327). These findings support that genetic factors contribute to PD susceptibility, while also emphasizing the need for larger multi-ethnic cohorts to identify additional risk loci, which could enhance the predictive performance of polygenic risk models. Integrating region-specific environmental exposures with genetic factors may further improve the accuracy of risk prediction in Asian populations.

Genome-wide cox regression analysis for PDD

Among the 4381 PD patients, 333 developed PDD. To assess the impact of variants contributing to the progression from PD to PDD, we conducted a two-stage genome-wide survival study (GWSS) among these patients (mean age: 78.74 ± 7.47 years, 49.25% men) and 1998 age- and sex-matched PD patients without dementia (mean age: 78.19 ± 7.51 years, 51.90% men).

We performed logistic regression analysis with a 1:6 age- and sex-matched case–control design, with adjustment for age, sex, disease duration, diabetes, hypertension, stroke history, and PC1–PC5. No variants reached genome-wide significance (P < 5 × 10⁻⁸); however, several loci showed suggestive associations (P < 5 × 10⁻⁶) (Supplementary Table S3). To further determine whether PD patients carrying these loci had a higher risk of PDD development compared with non-carriers, we used Cox proportional hazards models adjusted for age, sex, diabetes, hypertension, stroke history, and PC1–PC5. This analysis revealed three PDD risk loci that achieved genome-wide significance in the discovery cohort (participants from hospitals in northern Taiwan; 211 PDD cases and 1019 covariate-matched PD patients without dementia). These findings were confirmed in the replication cohort (participants from hospitals in southern Taiwan; 122 PDD cases and 979 covariate-matched PD patients without dementia) (Table 4).

Table 4.

Genetic variants associated with progression from PD to PDD in discovery, replication and combined cohorts.

Chr	Position	SNP	Gene	Minor allele	MAF	HR	95% CI	P value in discovery cohort	P value in replication cohort	P value in combined cohort
4a	77,835,861	rs10518182	CNOT6L, MRPL1	G	0.012	3.85	2.38–6.24	6.69E-6	6.37E-3	4.23E-08
4a	155,285,482	rs369720737	NPY2R, MAP9	A	0.013	3.48	3.00–5.45	6.12E-6	4.32E-3	4.27E-08
21b	41,815,851	rs141772267	PRDM15	T	0.012	4.20	2.67–6.60	1.34E-7	1.53E-4	5.41E-10

Chr, chromosome; SNP, single nucleotide polymorphism; MAF, minor allele frequency; HR, hazard ratio from the combined analysis; CI, confidence interval; P values from Cox proportional hazards models.

^aTwo variants were imputed.

^bOne variant was genotyped.

The lead variant rs141772267 in PRDM15, which encodes a zinc finger transcription factor involved in neurodevelopment,²³ was associated with a markedly higher risk of PDD (hazard ratio [HR] 4.20, 95% CI 2.67–6.60, P = 5.41 × 10⁻¹⁰ under the additive model; HR 4.01, 95% CI 2.52–6.39, P = 4.99 × 10⁻⁹ under the dominant model) (Supplementary Table S4, Fig. 3A). Two additional significant signals involved intergenic variants, rs10518182 (between CNOT6L and MRPL1) and rs369720737 (between NPY2R and MAP9) (Fig. 3A). Regional association plots are presented in the Supplementary Fig. S3A–C. In the Cox proportional hazards models, after adjustment for covariates in our cohort, PDD risk was no longer significantly associated with the APOE ε4 variant (rs199928040; HR: 1.99, 95% CI 1.30–3·06, P = 1.67 × 10⁻³) or the GBA1 variant (rs200282998; HR: 2.49, 95% CI 1.50–4.14, P = 4.41 × 10⁻⁴).

Fig. 3.

Genome-wide survival study (GWSS) of Parkinson's disease dementia (PDD). (A) Manhattan plot showing GWSS results for PDD in the TPMI dataset. The horizontal axis shows the chromosomal position, and the vertical axis shows the significance of tested markers from 6,660,482 SNPs in 333 PDD cases and 1998 PD without dementia controls. The gray dashed line denotes the genome-wide significance threshold (P < 5 × 10⁻⁸). (B–D) Kaplan–Meier survival curves in the TPMI cohort showing that PD patients carrying at least one minor allele of rs141772267 in PRDM15 (B), rs10518182 (C), or rs369720737 (D) exhibited significantly faster progression to PDD compared to non-carriers. (E–G) In the integrated cohort combining the TPMI and ethnicity-matched replication cohorts, the associations between each of the three variants and the risk of progression to PDD in PD patients were further strengthened: (E) PRDM15 rs141772267, (F) rs10518182, and (G) rs369720737. (H, I) Kaplan–Meier survival analysis showing that PD patients carrying two or more minor alleles across the three identified loci exhibited significantly accelerated progression to PDD compared to those with one or no minor allele. This association was observed in the combined TPMI cohort (H) and was further strengthened in the integrated joint cohort combining the TPMI and ethnicity-matched replication cohorts (I).

Next, Kaplan–Meier analyses revealed that PD patients carrying at least one risk allele of rs141772267 in PRDM15 exhibited faster cognitive decline, compared to non-carriers (CT or TT vs. CC, P = 4.99 × 10⁻⁹, Fig. 3B). Similarly, participants harboring risk alleles of rs10518182 or rs369720737 had significantly higher rates of progression to PDD during follow-up (P = 3.36 × 10⁻⁷ and P = 4.26 × 10⁻⁸, respectively; Fig. 3C and D). When these three variants were considered jointly, carrying at least two minor alleles compared to one or no minor allele conferred even faster progression to PDD (P < 2 × 10⁻¹⁶, Fig. 3H). Based on the sample size of the GWSS, the power was ∼80% and a HR of ≥3 would detect key susceptibility loci with a disease allele frequency of at least 3% (Supplementary Fig. S4). As the number of progression events to PDD was limited, statistical power was constrained. Therefore, variants reaching P < 5 × 10⁻⁶ were reported as exploratory signals to guide hypothesis generation, whereas genome-wide significance (P < 5 × 10⁻⁸) was reserved for loci supported by survival analysis and independent replication.

To validate these findings, we analyzed WGS data from an independent ethnicity-matched cohort of 871 PD patients. We excluded duplicates and first-, second-, and third-degree relatives, as identified through cryptic relatedness analysis, as well as individuals with <5 years of follow-up. Thirty-five patients progressed to PDD (mean age at conversion: 78.62 ± 5.12 years; 54.29% male) over a mean follow-up of 10.82 ± 3.67 years (Fig. 1A). Applying the same 1:6 ratio of cases to age- and sex-matched controls, we identified 210 PD patients without cognitive decline during follow-up (mean age: 76.98 ± 4.87 years; 51.43% male) for subsequent GWSS analysis (Fig. 1A, Supplementary Table S5). Consistent with the findings in the TPMI cohort, carriers of the risk allele of rs141772267 in PRDM15 exhibited a significantly increased risk of PDD development compared to non-carriers (P = 9.31 × 10⁻⁴, Supplementary Table S6). A forest plot was constructed to illustrate the effect estimates of the significant rs141772267 variant in PRDM15 across the discovery, replication, and independent WGS cohorts, in terms of PDD (Supplementary Fig. S5). The direction and magnitude of the effects were consistent across datasets, supporting the robustness of this association. Similarly, after adjustment for age, sex, and comorbidities (diabetes mellitus, hypertension, and stroke history), we observed elevated PDD risk during follow-up among participants carrying the minor alleles of rs10518182 (P = 4.57 × 10⁻³) or rs369720737 (P = 0.3848) (Supplementary Table S6).

Collectively, these findings demonstrated that the PRDM15 variant (rs141772267) was directly genotyped on the TPMI array (Probe ID: Affx-67962414), rather than imputed. The probe was validated via stringent quality control during array design and genotyping, showing a call rate of >99% and Hardy–Weinberg equilibrium P value of >0.05 in controls.^14,15 Among 1496 WGS samples from Taiwan Biobank, 1460 individuals were also genotyped using the TWB2.0 SNP array, revealing high concordance between platforms with discrepancies in only three cases.^14,15 A subsequent LD analysis across the PRDM15 region identified 1747 variants, only two of which exhibited moderate correlation with rs141772267 (r² > 0.2), while the remaining variants showed a mean r² of 0.0021 (SD = 0.0071). These results confirmed that this represents an independent signal. Moreover, the reliability of this association is supported by the replication in the WGS validation cohort.

When the ethnicity-matched WGS replication cohort was combined with the TPMI cohort to form a jointly integrated cohort, the associations of all three genetic variants with PDD risk were further strengthened. Carriers of the PRDM15 rs141772267 minor allele exhibited a markedly increased PDD risk (HR = 4.11, 95% CI 2.68–6.32, P = 1.03 × 10⁻¹⁰). Similarly, PDD risk was significantly elevated in carriers of the minor alleles of rs10518182 (HR: 3.84, 95% CI 2.46–5.97, P = 2.61 × 10⁻⁹) and rs369720737 (HR: 3.53, 95% CI 2.29–5.46, P = 1.36 × 10⁻⁸) (Supplementary Table S6).

Kaplan–Meier survival analyses further revealed that, within the jointly integrated cohort, carriers of the PRDM15 rs141772267 variant had a significantly higher PDD risk compared to non-carriers (P = 3.30 × 10⁻¹⁰, Fig. 3E), showing a stronger association than observed in the TPMI cohort alone (P = 4.99 × 10⁻⁹, Fig. 3B). PDD risk in the integrated cohort was also significantly increased among carriers of the minor alleles of rs10518182 (P = 7·24 × 10⁻⁹) and rs369720737 (P = 2.24 × 10⁻⁹) (Fig. 3F and G), compared to in the TPMI cohort alone (P = 3.36 × 10⁻⁷ and P = 4.26 × 10⁻⁸, respectively) (Fig. 3C and D).

Gene expression and interactome analysis

To gain biological insights regarding the novel PD risk loci identified in the merged Asian dataset, we examined gene expression using genotype-tissue expression data. Notable loci included intergenic and intronic variants near ELMO1, TECPR1, BORCS7, and HIP1R showed high expression in brain regions relevant to PD, including the caudate, putamen, frontal cortex, and substantia nigra (Fig. 4A).

Fig. 4.

Gene expression analysis of genes associated with Parkinson's disease (PD) or PD dementia (PDD). (A) Expression pattern of AOAH, ELMO1, TECPR1, BORCS7, and HIP1R. (B) Expression pattern of PRDM15, MRPL1, CNOT6L, NPY2R, and MAP9. (C) Analysis of the protein interaction network revealed clusters of proteins that interact with PRDM15.

PRDM15, the most significant candidate loci for PDD risk, was predominantly expressed in the cerebellum (Fig. 4B). NPY2R and MAP9 were highly expressed in cognition-related areas, such as the hippocampus, amygdala, and nucleus accumbens. While CNOT6L expression was limited in supratentorial regions, it was present in the cerebellum, which has emerged relevance to cognitive processes. PRDM15 is a key transcriptional regulator involved in neurodevelopment.²³ Interactome analysis reveals interactions with NFATC1 and NFATC2, both of which are glial-expressed mediators of neuroinflammation, and with serine/arginine-rich splicing factors (SRSFs), whose dysfunction contributes to T-cell-driven neuroinflammation and altered MAPT exon 10 splicing linked to tau pathology (Fig. 4C).²⁴ Additionally, PRDM15 interacts with DNA repair and RNA biogenesis factors, including DGCR8, recently implicated in PD-to-PDD progression.²³ These findings suggest a role for PRDM15 in neuroinflammation, tau co-pathology, and neurodegeneration in PDD.

Discussion

This study provides a comprehensive genetic analysis of PD and its progression to PDD in the Taiwanese population. The finding that key SNPs in SNCA and LRRK2 were associated with PD is consistent with previous studies,6, 7, 8, 9, 10, 11^,22 indicating shared genetic susceptibility across ethnicities. Furthermore, the merged Asian meta-analysis revealed additional PD risk loci, and a PRS incorporating these SNPs and clinical features effectively differentiated PD patients from controls. A novel PDD risk locus in PRDM15 conferred an over 4-fold risk of PDD, alongside two intergenic variants (CNOT6L-MRPL1 and NPY2R-MAP9). Moreover, carriers of two or more minor alleles exhibited a synergistically elevated risk of PDD development, and this finding was replicated in an independent Taiwanese cohort. Overall, our results demonstrate both shared and ethnicity-specific genetic factors contributing to PD and its progression to PDD.

Across ancestries, PD GWASs consistently converge on a small set of biological pathways, even when the lead SNPs and implicated genes differ among populations.6, 7, 8, 9, 10, 11^,22 Our results reinforced this pattern, and identified several relevant genes that are shared by European, Ashkenazi Jewish, and African populations, including SNCA, LRRK2, and TMEM175. Moreover, our findings are consistent with a recent WGS-based GWAS in a large Chinese cohort encompassing 1972 PD cases and 2478 controls, with replication in 8209 cases and 9454 controls, which demonstrated that the LRRK2 p.G2385R variant is an Asian-specific genetic risk factor for PD and also identified risk variants in SNCA and VPS13C.¹² Synaptic vesicle and α-synuclein biology are supported by SNCA and regulators of vesicle trafficking (eg, LRRK2 and HIP1R), which have been identified in the current study, prior GWASs, and meta-analyses.^7,22 TMEM175 is one of the most consistently replicated genetic risk loci for PD, across multiple ancestries. TMEM175 mediates lysosomal K⁺ conductance (which is critical for maintaining lysosomal membrane potential, luminal pH stability, and efficient autophagosome–lysosome fusion), and its biological properties provide a direct mechanistic link between lysosomal function, autophagy, and α-synuclein homeostasis.²⁵ Furthermore, the endo-lysosomal/autophagy pathway, supported by GBA1, has been highlighted in an African-specific GWAS,⁹ underscoring the convergence of PD genetic risk on vesicular and lysosomal mechanisms. The identification of these convergent genetic loci associated with PD warrants functional validation and offers insights into disease mechanisms and potential targets for mechanism-based therapy.

In addition to the shared loci, our combined Asian meta-analysis revealed several novel variants as potentially associated with PD risk, which may reflect population-specific genetic architecture. The intergenic variant rs602457, located between HLA-DRB1 and HLA-DQA1, demonstrated a suggestive association (P = 9.67 × 10⁻⁷). Although rs602457 did not reach the conventional threshold for genome-wide significance, its potential relevance is emphasized by its location within the highly immunologically active HLA region, and its prior implication in PD risk among European populations.⁷ HLA molecules play essential roles in antigen presentation and immune regulation. They can bind α-synuclein and activate T cells, with the magnitude of this immune response likely modulated by allelic variation.²⁶ Supporting the importance of this region, a recent trans-ethnic fine-mapping study identified a strong association between PD and amino acid position 13 in HLA-DRβ1, particularly in European populations.²⁷ Our findings thus extend the involvement of HLA-DR loci to Asian populations and support the role of immune-mediated mechanisms in PD pathogenesis. This aligns with previous reports of abundant HLA-DR-expressing microglia in the substantia nigra, highlighting the role of neuroinflammatory mechanisms in PD. ELMO1, a pattern recognition receptor, plays a role in microbial sensing and inflammatory responses in inflammatory bowel disease.²⁸ Notably, inflammatory bowel disease and PD share LRRK2 as a common risk gene.²⁹ ELMO1 interacts with NOD2,²⁸ whereas LRRK2 enhances NOD2 signaling to promote cytokine induction. This interplay may contribute to PD pathogenesis through inflammation amplification. TECPR1 is involved in autophagy regulation and was recently shown to be involved in the PD pathophysiology.³⁰ Deletion of TECPR1 causes accumulation of autophagosomes, whereas overexpression of TECPR1 in neurons reduces the neurotoxic protein aggregates in an autophagy-dependent manner.³⁰ BORCS7 is a known genetic risk factor for schizophrenia.³¹ BORCS7 encodes a subunit of the BLOC-1-related complex, a cytosolic protein complex critical for lysosomal positioning and trafficking, which processes had been linked to PD. HIP1R was recently identified to be a novel PD risk loci in a multi-ancestry GWAS.¹⁰ HIP1R is involved in vesicle trafficking, endocytosis, and membrane dynamics.¹⁰ The intronic variant rs10847864 is located in a transcription start site that is active in substantia nigra tissue.¹⁰ The identification of these PD-associated genetic loci in Asian populations highlights the potential involvement of immune-related pathways in PD pathogenesis, warranting further functional validation and offering insights into underlying disease mechanisms. However, we did not detect PD-associated loci near HEATR6, although an intronic variant in this gene was recently reported as a novel risk locus in a large Chinese population.¹² Future studies are warranted to investigate PD genetics across populations from different regions of Asia.

Novel genetic variants associated with PDD included an intronic variant in PRDM15 and two intergenic variants, one between CNOT6L and MRPL1 and the other between NPY2R and MAP9. Notably, individuals harboring at least two minor alleles experienced significantly faster cognitive decline to PDD, compared to those with one or no minor alleles. While PRDM15 has been proposed as a potential candidate locus for PD risk in a mixed Caucasian population,³² the reported association (P = 5 × 10⁻⁶) did not meet the conventional genome-wide significance threshold. Although the PRDM15 variant is of low frequency, its effect size indicates potential relevance for a subset of patients with PD who are at increased risk for cognitive decline. According to data from the Genome Aggregation Database (gnomAD), the rs141772267 variant in PRDM15 exhibits a relatively higher allele frequency in East Asian populations (MAF = 0.01842), while it remains rare in other populations, including South Asians (MAF = 0.0004942), Europeans (MAF = 0.0001823), and Africans (MAF = 0.0007462). Our findings demonstrate that a directly genotyped variant in PRDM15, rather than an imputed signal, is significantly associated with PDD risk. Notably, this variant was also confirmed in the WGS-analyzed validation cohort. These results offer novel insights into the genetic mechanisms underlying cognitive decline among individuals of East Asian ancestry with PD. More importantly, the implicated biological pathways, including tau splicing and neuroinflammation, offer insights into PDD pathogenesis that may be applicable beyond the variant carriers. Several large-scale GWASs in European populations have identified genetic variations linked to cognitive decline among PD patients.^3,4 A longitudinal survival study of 3821 PD patients⁴ pinpointed three variants: rs182987047 in RIMS2 and rs138073281 in TMEM108, which were rare in Asians; and rs8050111 in WWOX, which showed no significance in our dataset. The PRDM15 rs141772267 variant was not genotyped in a recent Korean cohort of 1029 PD patients, 347 of whom progressed to PDD.³³ A nearby SNP, rs2839398, was associated with a modestly increased risk of PDD (OR: 1.287, 95% CI 1.039–1.593, P = 0.02); however, it was only weakly correlated with rs141772267 (r² = 0.0063), and showed no association in our cohort (P = 0.3695). This signal should be interpreted cautiously, given the differences in LD structure and genotyping coverage. In another GWAS in a Chinese PD cohort (n = 450), rs75819919 in DPP6 was linked to accelerated cognitive decline⁵; however, this variant was not observed in our present cohort of PD patients. These findings suggest ancestral disparities in the genetic background, and that results may be influenced by sample size. PRDM15 regulates Notch signaling, which is crucial for forebrain development.²³ In vivo models with Notch mutations or RNA-mediated inactivation present neurocognitive deficits, particularly in long-term memory formation.³⁴ These findings suggest that variants in PRDM15 may disrupt Notch signaling, accelerating cognitive decline through altered postnatal neurodevelopment. The interactome analysis revealed that PRDM15 potentially interacts with gene clusters involved in neuroinflammation (NFACT1, NFATC2, PTEN, and PDK1), tau isoform regulation (SRSF family), DNA repair (XRC family), and RNA biogenesis (Ilf3, PRKRA, and DGCR8). Notably, PRDM15 may interact with the SRSF family of RNA-binding proteins, which regulate MAPT exon 10 transcription.²⁴ SRSF6 promotes exon 10 inclusion, leading to 4 R-tau expression, whereas SRSF4 inhibits exon 10 inclusion. Abnormal SRSF phosphorylation results in dysregulated tau exon 10 splicing, which leads to 3R/4 R-tau imbalance, exacerbating neuronal tau aggregation and accelerating cognitive decline in PD. We speculate that PRDM15 may contribute to PDD by regulating the SRSF protein family. Future studies should explore the PRDM15's downstream effectors, and its epigenetic role in regulating α-synuclein and tau homeostasis during PD-to-PDD progression.

In addition to PRDM15, we identified two intergenic variants associated with increased risk of cognitive decline among patients with PD: one located between CNOT6L and MRPL1, and the other between NPY2R and MAP9. CNOT6L encodes a catalytic subunit of the CCR4-NOT complex, a major regulator of mRNA deadenylation and decay, which plays a critical role in post-transcriptional gene regulation and protein quality control.³⁵ Dysregulation of this pathway is implicated in impaired neuronal homeostasis and synaptic plasticity, and is associated with intellectual disability, motor dysfunction, and speech delay.³⁵ MRPL1 encodes a mitochondrial ribosomal protein that is essential for mitochondrial protein synthesis. Dysfunction of MRPL1 compromises oxidative phosphorylation, increases reactive oxygen species production, and contributes to neuronal energy deficits and neurodegeneration. Mice exposed to chronic stress exhibit reduced MRPL1 expression in the hippocampus, a brain region particularly vulnerable to cognitive decline.³⁶ These findings suggest that the intergenic variant between CNOT6L and MRPL1 may influence the risk of cognitive decline in PD by modulating the expression of genes involved in synaptic maintenance and mitochondrial function, inducing heightened neuronal vulnerability in cognition-related networks. The second risk locus is an intergenic variant between NPY2R and MAP9. NPY2R encodes a G-protein-coupled receptor for neuropeptide Y, which has established roles in regulating neuroinflammation and neuroprotection. Neuropeptide Y can suppress microglial activation via its receptor, the translocator protein, which is highly expressed in the striatum and substantia nigra in a 6-hydroxydopamine-induced rat model of PD. Additionally, neuropeptide Y signalling supports memory formation and cognitive performance in animal studies.³⁷ The second risk locus is an intergenic variant between NPY2R and MAP9. NPY2R encodes a G-protein-coupled receptor for neuropeptide Y, which has established roles in regulating neuroinflammation and neuroprotection. Neuropeptide Y can suppress microglial activation via its receptor, the translocator protein, which is highly expressed in the striatum and substantia nigra in a 6-hydroxydopamine-induced rat model of PD. Additionally, neuropeptide Y signalling supports memory formation and cognitive performance in animal studies.³⁷ MAP9 encodes a microtubule-associated protein known as ASter-Associated Protein (ASAP), which is essential for microtubule organization, cytoskeletal integrity, and intracellular trafficking.³⁸ Enriched MAP9 expression in cortical regions, including the hippocampus, supports its potential roles in maintaining axonal transport, synaptic integrity, and possibly regulating microtubule-associated tau dynamics. These findings further support the hypothesis that coexisting AD-type pathology may contribute to PDD development. The intergenic variant between NPY2R and MAP9 may function as a regulatory element (eg, enhancer or silencer) affecting the transcriptional regulation of one or both genes. Dysregulation at this locus could lead to reduced NPY2R expression, impairing neuropeptide Y-mediated neuroprotection, or altered MAP9 expression, compromising cytoskeletal dynamics and neuronal resilience. There remains a need for functional studies to validate the regulatory effects of these intergenic variants, and their impact on neurodegeneration. The heterogeneous pathology of PDD with overlapping AD features warrants follow-up studies combining biomarker analyses, particularly tau homeostasis. Finally, although APOE ε4 and GBA1 variants were not significantly associated with PDD risk in our cohort after adjustment for covariates, the direction of effect was consistent with previous studies.^39,40 Future large-scale cohort studies including a greater number of patients who progress to PDD are needed to further clarify their roles in PDD susceptibility.

Strengths of this study include its hospital-based two-stage design, and the integration of EMR-based diagnoses with data regarding PDD-specific medications and longitudinal follow-up. Importantly, we utilized a Taiwanese-specific genotyping array, overcoming the limitations of previous GWASs that have predominantly relied on commercial arrays enriched for variants common among individuals of European ancestry. This approach enabled characterization of the genetic architecture underlying PD risk, and its progression to PDD, in a largely underrepresented East Asian population. This study also has several limitations. First, the genome-wide survival analysis included a relatively small number of patients who progressed to PDD, which may have reduced the statistical power to detect variants with modest effect sizes and increased the possibility of false-positive findings. Replication in larger independent cohorts, using the MDS Task Force criteria for PDD,^41,42 is therefore needed to validate the association between PRDM15 and PDD and to enable cross-cohort comparability. Additionally, since our analysis was based on Asians array-derived SNP data, future studies should be conducted using whole-genome sequencing, ideally conducted within the framework of international consortia,¹² to capture rare intronic variants, structural variants, and copy number variations that may contribute to PDD risk.

In conclusion, the present results reinforce the roles of SNCA and LRRK2 variants in PD risk across ethnicities and identify additional loci through Asian meta-analysis. Survival analysis revealed PRDM15 as a novel risk locus for cognitive impairment in PD. There remains a need for longitudinal studies to further investigate shared and ethnicity-specific variants related to PD and cognitive decline.

Contributors

Study concept and design: Chin-Hsien Lin and Cathy Shen-Jang Fann.

Data acquisition: Chin-Hsien Lin, Chien-Ching Chang, Hung-Hsin Chen, Wan-Jia Lin, Chia-Lang Hsu, Rung-Juen Lin, Yi-Jen Guo, Ting-Chun Fang, Shinn-Zong Lin, Chih-Yang Huang, Shuu-Jiun Wang, Jen-Fan Hang, Sun-Wung Hsieh, Mei-Chuan Chou, Tu-Hsueh Yeh, Chaur-Jong Hu, Fu-Chi Yang, Hsin-An Chang, Tsong-Hai Lee, Meng-Han Tsai, Ming-Che Kuo, Jyh-Ming Liou, Ming-Shiang Wu, Kye Won Park, Sun Ju Chung, Eng-King Tan, and Cathy Shen-Jang Fann.

Analysis and interpretation of data: Chin-Hsien Lin, Chien-Ching Chang, Hung-Hsin Chen, Wan-Jia Lin and Cathy Shen-Jang Fann.

Access and verify the data: Chin-Hsien Lin, Chien-Ching Chang.

Drafting of the manuscript: Chin-Hsien Lin.

Critical revision of the manuscript for important intellectual content: Chin-Hsien Lin and Cathy Shen-Jang Fann.

Funding acquisition: Cathy Shen-Jang Fann.

Study supervision: Chin-Hsien Lin and Cathy Shen-Jang Fann.

Responsible for the decision to submit the manuscript: Chin-Hsien Lin and Cathy Shen-Jang Fann.

Final responsibility to submit: Chin-Hsien Lin and Cathy Shen-Jang Fann.

Data sharing statement

Raw genotype or phenotype data cannot be made available due to restrictions imposed by the ethics approval. GWAS summary statistics of PD have been provided to the NHGRI-EBI GWAS Catalog and the study accession number is GCP001473.

Declaration of interests

All authors declare no competing interests.