Genome-wide association and population-tailored polygenic risk for Parkinson’s disease in Taiwan Authors: Yung-Tsai Chu, Yu-An Su, Chin-Hsien Lin, Chun-Hwei Tai, Yih-Ru Wu, Chien-Tai Hong, Yu-Wei Chen, Mong-Hsun Tsai, John Hardy, Kin-Ying Mok, Ruey-Meei Wu, East Asian Parkinson Disease Genomics Consortium, Global Parkinson’s Genetics Program Published: 2026 ABSTRACT AbstractGenetic risk factors for Parkinson’s disease (PD) remain under-characterized in East Asian populations. We assembled a Taiwanese case–control cohort (2245 PD; 2147 controls), genotyped participants using the Illumina NeuroBooster Array, and imputed 7.6 million variants with the Taiwan Biobank reference panel. Logistic-regression GWAS identified genome-wide significant associations at SNCA and MCCC1, with the lead SNCA signal located in intron 4; conditional and joint analyses suggested an additional 5′ SNCA component. We observed suggestive associations at GCH1, PPARGC1A, and GALNT13. Locus-focused haplotype analyses refined the SNCA signal, delineating an East Asian–enriched risk haplotype, and confirmed risk effects of LRRK2 p.G2385R and p.R1628P, including evidence consistent with a gene-dosage effect. A European-derived PRS showed modest discrimination in our cohort (AUC 0.59), while incorporating East Asian and Taiwan-relevant variants improved performance (AUC 0.62). Together, these results define the genetic architecture of PD in Taiwan, highlight shared and population-enriched risk components, and support ancestry-aware PRS construction for improved risk stratification.Subject terms: Diseases, Genetics, Neurology, Neuroscience Introduction Parkinson’s disease (PD) is the second most common neurodegenerative disorder worldwide.1 It is characterized clinically by tremor, rigidity, bradykinesia, and postural instability, and neuropathologically by selective dopaminergic neuron loss in the substantia nigra with Lewy bodies composed primarily of α-synuclein. Although a minority of PD cases arise from rare, high-penetrance variants, most cases are polygenic and multifactorial. Common variants of modest effect collectively account for a meaningful portion of PD heritability; the largest European-ancestry meta-GWAS to date identified 90 independent risk signals explaining an estimated 16–36% of the heritable risk.2 Yet the field’s understanding of PD genetics remains disproportionately derived from European cohorts, with fewer large-scale studies in East Asian and Taiwanese populations.3 Both clinical features and genetic risk architectures show ancestry-related differences.4 Asian patients, for example, have lower reported rates of motor complications such as dyskinesia than European patients, and several PD risk alleles differ in frequency and effect size across populations. Notably, the LRRK2 p.G2019S founder variant is enriched in European and certain West Asian groups, whereas LRRK2 p.G2385R and p.R1628P are established risk alleles in individuals of East Asian ancestry.5–7 At SNCA, ancestry-specific patterns have also been described: association signals and local haplotype structure vary by population8,9, and distinct rs356182-centered haplotypes show different relationships to risk in Latino and European cohorts, highlighting locus complexity relevant to non-European populations as well. In GBA1, recent GWAS in African cohorts have likewise identified novel PD risk variants among individuals of African ancestry.10 These differences have direct implications for risk prediction11. Cross-ancestry evaluations consistently show that polygenic risk scores (PRS) trained on European data underperform in non-European populations, arguing for population-tailored models. In response to this imbalance, the East Asian Parkinson Disease Genomics Consortium (EAPDGC) was formed to coordinate multi-center efforts across the region.12 We undertook the first Taiwan-wide PD GWAS to (i) map PD risk loci and related haplotypes in a Taiwanese cohort, (ii) benchmark effect sizes against European and pan-East Asian studies, and (iii) construct a Taiwan-optimized PRS to assess improvements in predictive performance over European- and Asian-derived baselines. Results Cohort demographics After quality control, the final cohort comprised 2251 individuals with Parkinson’s disease (PD) and 2268 controls (Supplementary Table S1 lists recruitment by site). Cases were slightly older at examination than controls (65.4 ± 11.2 vs. 63.5 ± 8.7 years) and more often male (51.5% vs. 44.4%). The mean age at PD onset was 60.5 ± 11.7 years. Key characteristics are summarized in Table 1. Age and sex were included as covariates in all association models. Population structure was evaluated by principal component analysis (PCA) using the 1000 Genomes Phase 3 reference panel; study participants clustered with the East Asian (EAS) superpopulation (CHB/CHS/JPT/KHV/CDX) (Supplementary Fig. S1). Case–control and site-level PCA were performed. We observed no obvious clustering or separation by disease status or by recruiting site, suggesting minimal population stratification and limited site-specific batch effects (Supplementary Fig. S2). Table 1Demographics of the study cohort after quality controlPD case, N = 2251Control, N = 2268Male Sex1184/ 2251 (51.5%)1007 /2268 (44.4%)Age at Exam65.4 (11.2)63.5 (8.7)Age at onset60.5 (11.7)NADisease Duration4.8 (4.4)NA1 Mean (SD); n / N (%)Summary of age, sex, and case–control distribution of participants included in the genome-wide association study after quality control. The detailed number of participants contributed by each recruitment site is provided in Supplementary Table 1. Demographics of the study cohort after quality control Summary of age, sex, and case–control distribution of participants included in the genome-wide association study after quality control. The detailed number of participants contributed by each recruitment site is provided in Supplementary Table 1. Genome-wide association analysis (Fig. 1; Tables 2–3) Genome-wide association results for Parkinson’s disease in the Taiwanese cohort. The Manhattan plot displays –log10p values for approximately 7.6 million variants after quality control and imputation. The red dashed line denotes the threshold for genome-wide significance (p = 5 × 10–8), and the blue dashed line indicates the threshold for suggestive significance (p = 1 × 10–5). Established PD risk loci, including SNCA, MCCC1, and GCH1, were successfully replicated. Additional signals in PPARGC1A and GALNT13 emerged as potential novel associations specific to the Taiwanese population. Although not exceeding the suggestive threshold, loci in VPS13C, LRRK2 (p.G2385R), and SCARB2 also demonstrated supportive evidence consistent with prior reports, suggesting potential relevance in East Asian populations. Table 2Top genetic risk loci associated with PD in TaiwanChrPosition (hg38)SNP (rsID)Nearest GeneEffect AlleleOther AlleleMAFOR_CIβSEPAnnotation489759478rs5860181SNCAAATGCATATT0.430.75 (0.69–0.81)–0.120.043.70e–11Most significat in Taiwan489765968rs3775427SNCAGA0.430.75 (0.69–0.81)–0.120.046.90e–11Most significat in Taiwan489761323rs6826785SNCATC0.430.76 (0.70–0.82)–0.120.041.60e–10Foo3183028420rs34822404MCCC1CT0.431.31 (1.21–1.42)0.120.047.20e–10Most significat in Taiwan3183042285rs10513789MCCC1TG0.431.3 (1.20–1.41)0.120.041.40e–09Nalls489836121rs3756059SNCAGA0.120.67(0.58- 0.76)0.400.071.86e–09Reported in the GWAS of RBD, located in 5’UTR489704960rs356182SNCAAG0.30.78 (0.71–0.86)–0.110.051.90e–07Nalls1454885803rs3783640GCH1TC0.461.23 (1.14–1.33)0.090.047.70e–07Taiwan1454882151rs11158026GCH1CT0.461.23 (1.14–1.33)0.090.041.40e–06Nalls424009221rs609501PPARGC1AAG0.481.22 (1.13–1.32)0.090.043.50e–06New2154088802rs10186542GALNT13AG0.121.35 (1.18–1.55)0.130.075.30e–06New Top genetic risk loci associated with PD in Taiwan Table 3Replication of established Parkinson’s disease loci and targeted look-up in the Taiwanese cohortChrPosition (hg38)rsID/variantGeneEffect AlleleOther AlleleMAFORBetaSEP1561994009rs12900645VPS13CGA0.3520.82 (0.76–0.90)–0.080.042.57E–05476217199rs76591264SCARB2/ FAM47EGC0.3291.20 (1.10-1.32)0.080.056.92E–051240363526rs34778348 (p.G2385R)LRRK2AG0.0331.59 (1.25–2.01)0.200.121.47E–04476213633rs11547135SCARB2/ FAM47ECT0.2950.846 (0.772–0.927)-0.170.053.55E–041240320043rs33949390 (p.R1628P)LRRK2CG0.0291.49 (1.16–1.92)0.170.131.94E–03 Replication of established Parkinson’s disease loci and targeted look-up in the Taiwanese cohort For the GWAS, after imputation and model-covariate completeness filters, 2245 PD cases and 2147 controls of Taiwanese ancestry remained, contributing ~7.6 million high-quality variants. The quantile–quantile plot showed minimal test-statistic inflation (λGC = 1.02), indicating adequate control of population structure (Supplementary Fig. S2). The Manhattan plot (Fig. 1) highlighted signals at SNCA and MCCC1, with additional support at GCH1 and several suggestive loci (P < 1×10–5). The most significant associations localized to established PD risk regions—SNCA, MCCC1, and GCH1. Beyond these, we observed suggestive associations at PPARGC1A and GALNT13, nominating potentially Taiwan-specific risk variants. Top results are summarized in Table 2 and Supplementary Table S2. Detailed regional association plots are provided in Supplementary Fig. S3. The lead association was rs5860181 within intron 4 of SNCA (odds ratio [OR] = 0.75, 95% CI 0.68–0.84, P = 3.7 × 10–11) (Fig. 2). Notably, the 5′ UTR variant rs3756059, previously reported in RBD13 and PD with RBD14, also reached genome-wide significance (Table 2). A prominent signal was detected at MCCC1 (rs34822404; OR = 1.31, 95% CI 1.19–1.44, P = 7.2 × 10–10), consistent with prior East Asian reports. At GCH1, we replicated association at rs3783640 (OR = 1.23, 95% CI 1.09–1.39, P = 7.7 × 10–7), which is in linkage disequilibrium with the previously reported rs11158026 signal. (Supplementary Fig. S4) SNCA regional association, site-level meta-analysis, conditional analysis, and colocalization. a–d LocusZoom-style regional association plots show −log₁₀(P) versus genomic position across ±300 kb of the indicated index SNP. Linkage disequilibrium (LD; r²) is calculated in cohort controls; the index SNP is shown in purple and other variants are colored by r² with the index. a SNCA regional association plot indexed on the Taiwan lead variant rs5860181 (chr4:89,759,478). The European index variant rs356182 is also labeled. The 5′ UTR variant rs3756059 additionally reaches genome-wide significance. b Site-level GWAS followed by meta-analysis at the SNCA locus (index rs5860181). c Conditional analysis conditioning on rs356182, identifying an additional 5′ signal at rs3756059 (chr4:89,836,121; conditional β = −0.26, P = 6.85e − 5). d Conditional analysis conditioning on rs5860181, revealing a secondary 5′ signal at rs1442145 (OR = 0.82, P = 1.88e − 5). e–g Colocalization of SNCA GWAS and whole-blood eQTL signals. e LocusCompare plot of SNCA GWAS versus eQTLGen (whole blood) association signals (−log₁₀ P). Points are colored by SNP-level posterior probability of colocalization (SNP PP.H4) from R “coloc”. The eQTL lead variant rs356162 shows higher posterior support for a shared causal signal with the GWAS association (PP.H4 = 0.67). f, g Regional association plots across the same SNCA window for the GWAS and eQTLGen signals, illustrating concordant association peaks consistent with colocalization at this locus. Among suggestive loci, PPARGC1A (rs609501; OR = 1.22, 95% CI 1.08–1.38, P = 3.5 × 10–6) and GALNT13 (rs10186542; OR = 1.35, 95% CI 1.01–1.45, P = 5.3 × 10–6) have not, to our knowledge, been reported in prior PD GWAS and warrant follow-up. We also observed the expected positive effects for LRRK2 p.G2385R and p.R1628P (Table 3). Finally, variants within VPS13C and SCARB2/FAM47E that were not in high LD with previously reported index variants from Nalls et al. or Foo et al. (Supplementary Fig. S3e-f) suggest potentially Taiwan-specific signals. Of note, the protective VPS13C signal at rs12900645 (chr15:61,994,009; hg38) lies within the putatively protective haplotype p.R153H–p.I398I–p.I1132V–p.Q2376Q (chr15:61,920,582–62,023,836) reported by Rudakou et al.15 In Taiwanese controls, linkage disequilibrium between rs12900645 and p.I1132V was modest (r² ≈ 0.44), indicating only partial tagging of the Rudakou haplotype. Across the four coding/synonymous variants reported by Rudakou et al., point estimates were protective but not statistically significant (e.g., p.I1132V OR = 0.91, P = 0.08). These findings are consistent with a protective background spanning chr15:61,920,582–62,023,836; however, in the Taiwanese cohort the intronic rs12900645 is the lead marker of the protective effect in VPS13C. Site-level GWAS, meta-analysis, and sensitivity analysis (Fig. 2b, Supplementary Fig. S5, and Supplementary Table S3) Because an independent external Taiwanese PD cohort was unavailable, we performed site-stratified GWAS and meta-analysis across three strata: NTUH + TWB + Landseed, CGMH, and SHH + TMU biobank. (Recruitment site information is summarized in Supplementary Table S1). Site-level PCA showed no obvious clustering by recruitment site (Supplementary Fig. S2a–c). The major signals at SNCA and MCCC1 remained significant across all sites and site-level meta-analysis (Fig. 2b and Supplementary Fig. S6). Compared with the overall Taiwan GWAS, GCH1 and PPARGC1A were less significant after stratification, likely owing to reduced power. At VPS13C (rs12900645), the effect direction was consistent in NTUH + TWB + Landseed and CGMH, whereas SHH + TMU showed a non-significant opposite direction; however, the combined meta-analysis remained significant. In contrast, WWOX (rs9922439) showed stronger evidence in the site-level meta-analysis (OR = 1.25, p = 1.18 × 10–6) than in the overall Taiwan GWAS (OR = 1.21, p = 1.21 × 10–5). Leave-one-site(-stratum)-out and NTUH-only analyses further supported the robustness of most lead loci (Supplementary Table S3). SNCA and SCARB2/FAM47E conditional analysis (Fig. 2c, d and Supplementary Table S4). We further conducted conditional and joint analyses implemented in GCTA-COJO16. At the SNCA locus, the Taiwan intron 4 lead variant rs5860181 remained associated after conditioning on the European 3′ region variant rs356182 (conditional β = −0.13, P = 0.002). Under the same conditional model, we detected an additional 5′ SNCA signal at rs3756059 (chr4:89,836,121; conditional β = −0.26, P = 6.85e − 5), a variant previously reported to be associated with REM sleep behavior disorder (RBD)13. Conversely, conditioning on rs5860181 attenuated the rs356182 signal (conditional β = −0.03, P = 0.45), while another 5′ signal rs1442145 (chr4:89,815,772) showed residual evidence of association (conditional β = −0.20, P = 2.1e − 5). In joint modeling, rs5860181 remained significant when modeled together with rs356182 (rs5860181 P = 3.1e − 5; rs356182 P = 0.31). When jointly modeling four variants (rs5860181, rs356182, rs1442145, rs3756059), rs5860181 (P = 1.0e − 4) and rs1442145 (P = 0.007) remained significant. At the SCARB2/FAM47E locus, the European lead variant rs11547135 (chr4:76,213,633) recently reported by Sun et al.17 lies ~3.5 kb from our Taiwan lead variant rs76591264. Using GCTA-COJO two-variant joint modeling, rs76591264 remained more significant (joint β = 0.138, P_joint = 0.00759) than rs11547135 (joint β = −0.105, P_joint = 0.0436), supporting rs76591264 as the stronger independent association in our cohort. The two variants showed a moderate inverse correlation in the LD reference (r = −0.44), consistent with partial LD rather than complete collinearity. Genotype analyses of LRRK2 risk alleles (Table 4) Table 4Distribution of East Asian LRRK2 risk genotypes and associated PD oddsALRRK2 risk variant number01*2**LRRK2 carrier rateControl (n = 2268)204721659.7%PD (n = 2251)19352971914.0%Onset age60.75 ± 11.1959.17 ± 11.2562.00 ± 9.38One-way ANOVA: p value 0.07BOdds ratio95% CIp valuesingle variant vs. non-carrier1.491.23–1.794.1E–05double variants vs. non-carrier4.291.59–11.60.004double variants vs. single variant2.891.06–7.90.039Odds ratios from logistic regression adjusted for age, sex, and top 15 PCs; reference = noncarriers. “Double-variant” includes G2385R/G2385R, R1628P/R1628P, and G2385R/R1628P compound heterozygotes.* either LRRK2 G2385R or R1628P one variant carrier.** double variants : G2385R homozygous, R1628P homozygous or G2385R + R1628P compound heterozygous. Distribution of East Asian LRRK2 risk genotypes and associated PD odds Odds ratios from logistic regression adjusted for age, sex, and top 15 PCs; reference = noncarriers. “Double-variant” includes G2385R/G2385R, R1628P/R1628P, and G2385R/R1628P compound heterozygotes. * either LRRK2 G2385R or R1628P one variant carrier. ** double variants : G2385R homozygous, R1628P homozygous or G2385R + R1628P compound heterozygous. The combined carrier rate of East Asian LRRK2 risk variants was 9.7% in controls and 14.0% in PD. Relative to noncarriers, individuals carrying one variant (either p.G2385R or p.R1628P) had OR = 1.49 (95% CI 1.23–1.79; P = 4.1 × 10–5). Carriers of two risk variants (p.G2385R homozygotes, p.R1628P homozygotes, or p.G2385R/p.R1628P compound heterozygotes) had OR = 4.30 (95% CI 1.59–11.6; P = 0.004) versus noncarriers and OR = 2.89 (95% CI 1.06–7.91; P = 0.039) versus single-variant carriers, consistent with a gene-dosage effect. Age at onset (mean ± SD) did not differ significantly across groups: noncarriers 60.75 ± 11.19 years, one-variant carriers 59.17 ± 11.25 years, and two-variant carriers 62.00 ± 9.38 years. In an exploratory subgroup comparison, compound heterozygotes (p.G2385R/p.R1628P) showed a numerically earlier onset (55.6 ± 5.8 years) versus noncarriers (60.75 ± 11.19 years), a trend that did not reach significance (t test P = 0.06); this analysis is limited by small sample size. Haplotype structure at SNCA, LRRK2, and GCH1 (Fig. 3) Haplotype structure and association of SNCA and LRRK2 in the Taiwanese Parkinson’s disease cohort. a Linkage disequilibrium (LD) heatmap (r²) among five representative SNCA variants (rs2870004, rs356182, rs356203, rs6826785, rs3775427) computed in controls. b Case–control frequencies of common SNCA haplotypes (allele order as in panel a); bases are color-coded by nucleotide. Protective haplotypes AGTTG and AATTG are highlighted. c Forest plot of SNCA haplotype associations (odds ratio [OR] with 95% confidence interval [CI]; log scale) with corresponding P values; AGTTG and AATTG show reduced PD risk. d LD heatmap (r²) for five LRRK2 missense/region-specific variants—N551K (rs7308720), R1398H (rs7133914), R1628P (rs33949390), chr12:40341489, and G2385R (rs34778348)—illustrating an East Asian–patterned LD structure. e LRRK2 haplotype composition and case–control frequencies (allele order as in panel d). f Forest plot of adjusted LRRK2 haplotype associations (OR, 95% CI, P). Haplotypes CGGCA and CGCTG exhibit protective effects relative to the reference CGGTG. LD linkage disequilibrium, OR odds ratio, CI confidence interval. Analysis details: LD was estimated with PLINK v1.9 (–r²) in controls. Haplotypes were inferred by EM (PLINK) and tested for association using the R package haplo.stats, with models adjusted for age, sex, and the first five principal components. To probe ancestry-specific architecture, we performed regional haplotype analyses for SNCA, LRRK2, and GCH1. SNCA The lead variant rs3775427 (intron 4) resides on an East Asian–enriched haplotype distinct from the European rs356182–centered configuration. Using 1000 Genomes East Asian reference data via LDlink, rs3775427—the top SNCA signal in Taiwan—was tightly correlated with rs6826785 (D′ = 0.996, r² = 0.988) and only partially correlated with rs356182 (D′ = 0.83, r² = 0.469). Haplotype analysis of five representative variants (rs2870004, rs356182, rs356203, rs6826785, rs3775427) identified two common configurations, AGTTG and AATTG, that were associated with reduced PD risk (Fig. 3a–c). Notably, both protective haplotypes carry the rs3775427-G allele, whereas the rs356182 allele differs between them (A vs. G), consistent with rs3775427-G (or variants on its tight LD background) tagging the protective effect. These findings support population-specific LD blocks at SNCA, with the Taiwanese pattern aligning more closely with other East Asian cohorts than with Europeans. LRRK2 We evaluated five missense/region-specific variants—p.N551K (rs7308720), p.R1398H (rs7133914),p. R1628P (rs33949390), an intronic variant at chr12:40341489, and East Asian–specific p.G2385R (rs34778348). LD patterns indicated three largely independent variant groups (N551K–R1398H; R1628P; G2385R). Haplotype reconstruction followed by adjusted logistic regression identified CGGCA and CGCTG as risk haplotypes relative to the reference CGGTG, whereas the N551K–R1398H combination did not show a significant protective effect (Fig. 3d–f). Further principal component analysis (Fig. 4) shows that carriers of both LRRK2 Asian variants cluster mainly within CHS (Southern Han Chinese). p.R1628P carriers are additionally observed in CDX (Chinese Dai) and KHV (Kinh Vietnamese), suggesting a more southern extension, whereas p.G2385R carriers are also seen in CHB (Han Chinese in Beijing) and JPT (Japanese), indicating a more northeastern distribution. LRRK2 p.R1628P and p.G2385R on East Asian population PCA (PC2–PC3). Principal component analysis using LD-pruned autosomal SNPs (GRCh38) from the 1000 Genomes East Asian populations (CDX, CHB, CHS, JPT, KHV) and a Taiwanese cohort (TW). Points are colored by population; carriers of LRRK2 p.R1628P (left) and p.G2385R (right) are indicated by red hollow circles. Related individuals were removed prior to PCA. GCH1 In Taiwan, rs3783640 is in LD with rs11158026 reported previously; both are only in partial LD with rs841 (r² ≈ 0.31; Supplementary Fig. 2). Haplotype analyses indicated that configurations carrying rs3783640 and rs11158026 are associated with PD risk, whereas rs841 alone was not significantly associated, suggesting that prior protective signals attributed to rs841 may be mediated by rs11158026/rs3783640. Together, these analyses illustrate that while SNCA, LRRK2, and GCH1 are globally implicated in PD, internal LD and haplotypic configurations differ by ancestry, underscoring the value of population-tailored analyses. Colocalization analysis At SNCA, colocalization with whole-blood cis-eQTLs from eQTLGen supported a shared signal (PP.H4 = 0.67; 2627 SNPs) (Fig. 2e–g). The lead PD GWAS variant (rs5860181) and lead blood eQTL variant (rs356162) fell within the same LD block in the Taiwanese reference (D′ = 1; r² = 0.21), consistent with a shared regional association despite different lead variants. In metaBrain EAS cortex eQTLs, evidence was limited (PP.H4 = 0.08). For PPARGC1A (rs609501 region), colocalization was not supported in eQTLGen (PP.H4 = 0.08) or metaBrain EAS cortex, and nearby expressed genes also showed uniformly low support (PP.H4 < 0.1). At VPS13C, colocalization was not supported in whole blood (PP.H4 = 0.05) or metaBrain EAS cortex (PP.H4 = 0.04). Although the lead PD and blood eQTL variants were in the same LD block (D′ = 0.97; r² = 0.56), the pattern was more consistent with coincident LD-driven signals than a shared causal variant. At SCARB2/FAM47E, colocalization was not supported for SCARB2 in whole blood (PP.H4 = 0.014) or cortex (PP.H4 = 0.020). A nearby gene, THAP6, showed only weak blood evidence (PP.H4 = 0.121). Functional annotation of associated loci CADD-based annotation showed that both LRRK2 p.G2385R and p.R1628P had high CADD scores (23.200 and 26.400), supporting the functional impact. rs11547135 at the SCARB2/FAM47E locus also showed a potentially deleterious CADD score (10.780). To assess the potential regulatory relevance of the associated variants, we performed functional annotation using GTEx v10 and SCREEN (Search Candidate Regulatory Elements by ENCODE) (Supplementary Tables S7 and S7B). At the SNCA locus, multiple variants showed regulatory evidence. In GTEx, rs3756059, rs6826785, and rs356182 were associated with SNCA expression across several brain regions, including basal ganglia and cerebellar tissues. Consistently, SCREEN annotation showed that the Taiwan SNCA lead variant rs5860181 overlapped two distal enhancer elements, whereas the 5′ variant rs3756059 mapped to a promoter element, supporting regulatory complexity across this locus. Additional brain eQTL support was observed for GCH1 variants (rs11158026 and rs3783640) and for rs12900645, which was associated with VPS13C expression; notably, rs12900645 also overlapped a distal enhancer in SCREEN. rs34822404 at MCCC1 overlapped a CA–H3K4me3 regulatory element, whereas variants at MCCC1, GALNT13, and PPARGC1A showed predominantly peripheral-tissue eQTL association in GTEx. At the SCARB2/FAM47E locus, the Taiwan-associated variants rs76591264 mapped to distal enhancers, while nearby previously reported European-region variants rs11547135 overlapped promoter or proximal enhancer elements. No nearby cCREs were identified for GCH1 (rs3783640), PPARGC1A (rs609501), or GALNT13 (rs10186542). Overall, these annotations support a regulatory basis for several of the observed associated loci. Cross-ancestry comparison (Fig. 5; Supplementary Table S3) Replication of genome-wide significant variants from Nalls et al. (2019) and Foo et al. (2020). Scatter plots compare the effect sizes (log odds ratios) of PD-associated variants identified in previous GWAS with those observed in the Taiwanese cohort. For detailed statistics, please refer to Supplementary Table 4. Green points represent variants that were not statistically significant in the current dataset. A modest correlation was observed with the European meta-analysis by Nalls et al. (2019) (R = 0.67, p = 2.3 × 10⁻¹¹), reflecting limited cross-ancestry transferability of risk alleles. In contrast, effect sizes showed a strong positive correlation with the pan-Asian GWAS by Foo et al. (2020) (R = 0.96, p = 7.5 × 10⁻⁶), indicating a high degree of genetic concordance and shared PD risk architecture within East Asian populations. We compared effect sizes in this Taiwanese GWAS with those from a large European meta-analysis (Nalls 2019) and a pan-East Asian GWAS (Foo 2020). Scatter plots of log-odds ratios showed a modest but significant correlation with European estimates (Pearson’s r = 0.67, P = 2.3 × 10–11), with over half of variants not reaching significance in our cohort (green points in Fig. 3a). In contrast, effects were strongly concordant with East Asian estimates (r = 0.96, P = 7.5 × 10–6), reflecting shared risk architecture within East Asia. Notably, the MCCC1 risk allele exhibited a larger effect in the Taiwanese cohort (OR = 1.32) compared with European cohorts (OR ≈ 1.1), suggesting population-specific modulation of genetic risk. Polygenic risk scores and population-specific improvement (Fig. 6) Polygenic risk score (PRS) performance for Parkinson’s disease in the Taiwanese cohort. a Overview of the PRS training and evaluation workflow using recruitment-site holdout testing. b Site-separated out-of-sample PRS evaluation with CGMH held out as an independent test set (290 PD cases / 451 controls; ~16% of the cohort). PRS models were trained on the remaining sites and evaluated in CGMH: (i) Nalls (EUR-only), (ii) Nalls + Asia, (iii) Nalls + Asia + Taiwan (4 variants), and (iv) TW-16SNV (Asia + Taiwan + GCH1). In the held-out CGMH test set, AUCs (95% CI) were 0.59 (0.55–0.63) for PRS-EUR (76 SNVs), 0.61 (0.568–0.651) for PRS-EUR+Asia (87 SNVs), 0.62 (0.579–0.662) for PRS-EUR+Asia+Taiwan (91 SNVs), and 0.61 (0.567–0.650) for TW-16SNV (16 SNVs). We evaluated PRS models to test whether adding East Asian and Taiwan-relevant variants improves discrimination. Using published EUR weights (Nalls et al.), 76 of 90 variants were retained after harmonization and QC, and PRS-EUR achieved an AUC of 0.590 (95% CI, 0.574–0.607). Adding 11 East Asian lead variants (Foo et al., 2020) yielded PRS-EUR-Asia (87 variants) and improved AUC to 0.606 (95% CI, 0.590–0.623) (Supplementary Fig. S7). To avoid overfitting when incorporating Taiwan-specific variants, we used a site-separated out-of-sample design, holding out CGMH as an independent test set (290 cases/451 controls) and training on the remaining sites. We evaluated PRS-EUR, PRS-EUR-Asia, PRS-TW (91 variants) by adding four Taiwan-relevant variants (LRRK2 p.G2385R, LRRK2 p.R1628P, VPS13C rs12900645, SCARB2/FAM47E rs76591264), and PRS-TW-16 SNVs (11 East Asian + 4 Taiwan + GCH1). In the held-out CGMH set, AUCs (95% CI) were 0.59 (0.55–0.63), 0.61 (0.568–0.651), 0.62 (0.579–0.662), and 0.61 (0.567–0.650), respectively (Fig. 6b), indicating that a compact Taiwan-focused model can achieve discrimination comparable to larger EUR-derived PRS models. Discussion This is the first large-scale genome-wide association study (GWAS) of Parkinson’s disease (PD) in a Taiwanese cohort. It refines PD genetics in East Asia and underscores the value of ancestry-specific analyses. We replicated core PD loci (SNCA, MCCC1, GCH1) and East Asian LRRK2 risk alleles (p.G2385R and p.R1628P), supporting conserved biological pathways across ancestries.2,3,18 At the same time, we observed clear ancestry-patterned features at SNCA. The Taiwanese lead variants rs3775427 and rs5860181 (both intron 4) reside on an East Asian–enriched haplotype. LDlink analyses using 1000 Genomes East Asian data showed tight correlation with rs6826785 (D′ ≈ 0.996, r²≈0.988) but only partial correlation with the European lead rs356182 (D′ ≈ 0.83, r²≈0.47), consistent with pan-Asian GWAS architecture.2,3,9 Conditional analyses reinforced this pattern: the intron-4 variants rs5860181 remained associated when conditioning on rs356182, whereas the effect at rs356182 was attenuated when conditioning on rs5860181. We also identified evidence for a possible secondary 5′ SNCA signal, including rs3756059, a variant previously implicated in RBD and PD with RBD13,14. This observation raises the possibility that 5′ regulatory variation at SNCA may contribute not only to PD susceptibility but also to phenotypic heterogeneity. Notably, the LARGE-PD consortium, which performed a similar conditional analysis (conditioning on rs356182), did not observe any residual intron-4 or 5’ UTR signal19. This contrast implies that the intron-4 association may be more prominent in the Taiwanese and broader East Asian populations. Haplotype-based association identified two protective configurations (AGTTG, AATTG) that both carry rs3775427-G; notably, AGTTG retains rs356182-G on the canonical East Asian background. Together, these findings indicate that variation within the intron 4 region contributes meaningfully to disease risk, and that SNCA risk and protection in Taiwan are better understood in the context of regional haplotypes rather than single variants. The putative 5′ UTR SNCA signal will require validation in independent East Asian cohorts, ideally with phenotype-informed analyses including RBD. This has translational relevance as α-synuclein–targeted assays and therapies evolve, since underlying genetic polymorphisms may differentially influence α-synuclein expression, conformation, and treatment response.20 LRRK2 provides another example of ancestry-enriched PD risk architecture in East Asians. In our cohort, the two major Asian risk variants, p.R1628P and p.G2385R, were observed on distinct haplotypic backgrounds, consistent with previous reports that these alleles have different geographic distributions within East Asia5,21,22. By contrast, the commonly described East Asian protective haplotype p.N551K–p.R1398H was not significantly protective in our dataset23, underscoring possible heterogeneity even within Asian populations. Functionally, p.R1628P is located in the COR domain and has been reported to enhance LRRK2 interaction with Cdk57,24, leading to increased kinase activity, whereas p.G2385R, located in the WD40 domain25,26, appears to have more complex consequences, including altered domain structure, impaired vesicle-related interactions, and context-dependent effects on kinase activity. The observed gene-dosage effect may reflect the structural biology of LRRK2 rather than simply a cumulative statistical burden. Because LRRK2 forms dimers and the WD40 domain contributes to dimerization25, different combinations of wild-type and mutant subunits may generate distinct functional outputs. In this context, the higher PD risk in double-variant carriers27 may reflect enhanced kinase dysregulation in mutant-containing complexes. This interpretation is supported by recent functional data showing that double-variant carriers (p.G2385R + p.R1628P) have markedly increased pRab10Thr73 phosphorylation compared with controls, consistent with heightened LRRK2 pathway activation28. Taken together, these findings suggest that combined Asian LRRK2 risk variants may exert additive or synergistic biological effects and support the relevance of genotype-informed stratification in future mechanistic and therapeutic studies. Given the relative enrichment of double-variant carriers (0.8% of PD cases vs 0.2% of controls) in Taiwan, these individuals represent a priority group for prodromal PD research and preventive intervention. Notably, caffeine intake may modify PD risk among LRRK2 variant carriers in Asian populations29, and first-in-class type II kinase inhibitors that stabilize the inactive LRRK2 conformation have recently been reported30, highlighting opportunities for pharmacogenomic stratification and targeted prevention. At GCH1, we replicated association at rs3783640, which is in LD with the previously reported rs11158026 signal31; in contrast, rs841-A alone was not protective in our data, and haplotype analysis suggested that earlier protective observations attributed to rs841 are likely mediated by the correlation with rs11158026-T32. For MCCC1, the Taiwanese effect size (OR ≈ 1.32) exceeded European reports (≈1.1) and aligns with observations in southern Chinese cohorts.33 We also observed a suggestive association at PPARGC1A, a master regulator of mitochondrial biogenesis/oxidative stress previously linked to PD risk34 or age at onset35 and to essential tremor (ET)36; rare-variant data in Han Chinese further support involvement of PPARGC1A in PD34, motivating cross-disorder analyses in Taiwan. PPARGC1A is a biologically plausible therapeutic node (PGC-1α). In Taiwan’s National Health Insurance Database, exposure to the PPARγ agonist pioglitazone was associated with a lower incidence of PD (adjusted HR 0.66, 95% CI 0.57–0.78; dose-dependent)37, whereas a phase 2 trial in early PD did not demonstrate disease-modifying effects.38 These apparently discordant findings suggest context-dependent efficacy. We hypothesize that the pioglitazone–PD association is modified by PPARGC1A genotype, warranting genotype-stratified pharmacoepidemiology and prospective validation studies. Finally, signals within VPS13C and SCARB2 not in high LD with previously reported leads suggest potential Taiwan-specific haplotypes that warrant replication and conditional fine-mapping31. In VPS13C, we detected an intronic protective association at rs12900645 that lies within a previously described protective region (chr15:61,920,582–62,023,836) encompassing the p.R153H–p.I398I–p.I1132V–p.Q2376Q haplotype reported by Rudakou et al.15. Linkage disequilibrium between rs12900645 and p.I1132V was modest (r²≈0.44), and the four coding/synonymous variants showed protective but non-significant effects in our data, together suggesting a shared protective background with rs12900645 as the lead tag in Taiwanese individuals. Given that VPS13C encodes a lipid-transfer protein implicated in lysosomal–mitochondrial homeostasis, this signal is biologically plausible but requires confirmation. GALNT13, a glycosyltransferase, emerged as a putative novel locus, echoing prior Asian evidence at WBSCR17/GALNT173 and raising the possibility that O-glycosylation pathways contribute to PD susceptibility in Asian populations39. Recent studies have also shown that glycation of α-synuclein promotes its aggregation and exacerbates neuroinflammatory responses.40 Cross-ancestry comparisons showed effect sizes that correlated modestly with European meta-analysis and strongly with pan-East Asian estimates, reflecting both shared biology and ancestry-patterned architecture.2,3 This also helps explain limitations in PRS transferability. For LRRK2, the Nalls 90-variant panel includes European-informative variants such as p.G2019S (rs34637584)41, but this variant is extremely rare in East Asians and was absent or near-monomorphic in our dataset; in contrast, rs76904798 was retained because it is available in our data. At GBA1, several variants in the Nalls panel are rare or absent in East Asians and were not reliably represented in our imputed data, while the locus itself is difficult to evaluate by SNP-array imputation because of GBAP1 and complex local structure. Conversely, Taiwanese-relevant variants such as p.L444P/L483P42 are not well captured by standard imputation43. Therefore, GBA1 is not included in this study. Consistent with limited cross-ancestry PRS portability, the European-derived PRS achieved modest discrimination in our cohort (AUC = 0.590). Adding East Asian-associated variants improved the AUC to 0.606, and further inclusion of Taiwan-specific variants increased it to 0.620. These results show that incorporating ancestry-relevant variants improves risk prediction in Taiwanese PD. Importantly, a compact 16-SNV model based on Asia + Taiwan + GCH1 variants performed similarly to models that also included a larger set of European-derived variants, suggesting that a smaller, population-tailored panel may be sufficient for practical PRS application in Taiwanese and possibly other East Asian populations. Functional follow-up provided supportive evidence at the SNCA locus, where colocalization with eQTLGen44 whole-blood cis-eQTLs suggested a shared regulatory signal. However, interpretation of these analyses is limited by the lack of adequately powered, ancestry-matched brain molecular QTL resources. Although we incorporated the metaBrain East Asian cortex eQTL45 dataset to improve ancestry matching, its modest sample size and incomplete locus-level variant coverage reduced power for colocalization and likely limited detection of shared causal signals. In addition, bulk cortex data may not capture cell type–specific or region-specific regulatory effects relevant to PD. Accordingly, our functional analyses should be regarded as supportive and hypothesis-generating rather than definitive, and future studies using larger East Asian brain eQTL datasets or cohort-matched molecular QTL resources will be important to refine the regulatory interpretation of population-specific signals. Several limitations merit mention. First, we used SNP arrays with imputation rather than comprehensive sequencing, which reduces sensitivity to rare coding variants and structural variation and performs suboptimally at complex loci (e.g., GBA1 with its pseudogene)46. Deep whole-genome sequencing and/or long-read approaches with locus-specific validation would address this. Second, this is a Taiwan-focused cohort designed to fill a clear gap; findings may not capture the full heterogeneity across East and Southeast Asia. Replication in independent EAPDGC sites12 and additional regional cohorts will be important. Third, Taiwan-specific variants have not yet been externally replicated because a truly independent Taiwanese dataset for evaluating Taiwan-specific components is currently unavailable. Likewise, although we held out certain recruiting sites as internal validation sets, the PRS-TW model was developed and evaluated within the same overarching cohort, introducing the potential for overfitting and optimistic bias in estimated discrimination. Accordingly, the observed AUC gains—particularly those attributable to Taiwan-specific variants—should be interpreted as preliminary until validated in an independent Taiwanese cohort. Moreover, the PRS performance was assessed using logistic regression without clinical or lifestyle covariates (e.g., caffeine29, smoking), prodromal markers, or biomarkers. Future work should evaluate integrative models that combine genetics with environmental and clinical features, consider machine-learning approaches47 with rigorous internal/external validation, and report calibration and clinical utility (e.g., decision-curve analysis).48 In conclusion, this Taiwan-wide GWAS refines PD genetics in East Asians and underscores the value of ancestry-specific analysis. We confirmed established loci (SNCA, MCCC1, GCH1), defined an East Asian intron 4 SNCA haplotype underlying risk and protection, and showed that the common Asian LRRK2 variants p.G2385R and p.R1628P have a gene-dosage effect. We also identified a VPS13C protective region and suggestive associations at PPARGC1A and GALNT13, although these population-specific loci require external replication. A Taiwan-optimized PRS improved discrimination over a European-derived model, but because it was developed and tested within the same cohort, potential overfitting should be acknowledged and external validation remains necessary. These findings provide a basis for future replication in East and Southeast Asia, alongside fine-mapping, functional follow-up, and integration with clinical and environmental data. Methods Study population and cohort design Participants were recruited through the East Asian Parkinson Disease Genomics Consortium (EAPDGC) via a multi-center collaboration. Taiwanese recruitment sites included National Taiwan University Hospital (NTUH), Chang Gung Memorial Hospital (CGMH), Shuang Ho Hospital/Taipei Medical University (SHH/TMU), Landseed Hospital and Taiwan Biobank.49,50 Parkinson’s disease (PD) diagnosis followed the Movement Disorder Society (MDS) Clinical Diagnostic Criteria.51 We excluded atypical parkinsonism and secondary parkinsonism, including drug-induced parkinsonism, vascular parkinsonism, and normal-pressure hydrocephalus. Participants with major comorbid neurological or psychiatric conditions that could confound PD diagnosis or clinical assessment (e.g., prior stroke with significant sequelae, epilepsy, schizophrenia) were also excluded. All participants provided written informed consent. The study protocol was approved by the NTUH Institutional Review Board (NTUH-REC No. 201905035RINC). Genotyping and quality control Genotyping used the Illumina NeuroBooster Array, which combines the Global Diversity Array backbone with custom neuro-focused content.43 Quality control (QC) was performed in PLINK v1.9 and v2.0.52,53 Samples were excluded for low call rate (<95%), sex discordance, or excess relatedness (one of each pair with PI_HAT > 0.20 removed). Variants were removed for Hardy–Weinberg disequilibrium in controls (p < 1×10⁻⁴), missingness >5%, or minor allele frequency (MAF) < 0.005. Principal component analysis (PCA) was performed to control for population structure. Association models adjusted for age, sex, and ancestry PCs derived from study genotypes. Imputation Post-QC genotypes were imputed using the Taiwan Biobank (TWB) imputation service, which employs Minimac4 and a population-specific reference panel of ~1500 TWB whole-genome sequences.49,54 Imputation was performed on the TWB server pipeline; we applied post-imputation filters of MAF > 0.01 and Rsq>0.30, yielding ~7.6 million high-quality variants for analysis. Principal component analysis (PCA) Autosomal, biallelic SNPs were LD-pruned in PLINK v1.9 (parameters: --maf 0.10; --indep 50 5 1.5), and principal components were then computed on the pruned set using --pca. Study genotypes were merged with the 1000 Genomes Project Phase 3 (GRCh38) reference to anchor ancestry inference. Scatter plots of top PC were generated in R (v4.4.2) using ggplot2 to visualize clustering; participants aligned with the East Asian superpopulation (Fig. 3 and Supplementary Fig. S1). Statistical analysis Genome-wide association testing used logistic regression in PLINK 2.0 with PD status as the outcome, adjusting for age, sex, and the top 15 principal components. Genome-wide significance was defined as p < 5 × 10–8 and suggestive significance as p < 1 × 10–5. Analyses were later conducted in R (v4.4.2).55 Specifically, we ran logistic regression in PLINK 2.0 using the --condition flag to include the index variant as a covariate, adjusting for age, sex, and the first 15 principal components. Site-level GWAS, meta-analysis and sensitivity analysis We performed site-stratified GWAS across three recruitment strata (NTUH + TWB+Landseed, CGMH, and SHH + TMU) using logistic regression with covariate adjustment for age, sex, and ancestry principal components. Stratum-specific additive-model summary statistics were combined using PLINK 1.9 fixed-effects inverse-variance meta-analysis (plink19 --meta-analysis + study --out ), with between-stratum heterogeneity assessed by Cochran’s Q and I². Key loci were summarized using forest plots displaying per-allele odds ratios with 95% confidence intervals for each stratum and the combined estimate. (Supplementary Fig. 5 and Supplementary Table 4) Linkage disequilibrium and haplotype analysis Population LD context was evaluated using LDlink to query East Asian reference panels.¹² Within the Taiwanese cohort, pairwise LD (r²) was estimated with PLINK (–r2). Haplotype visualization for SNCA, LRRK2, and GCH1 used the geneHapR R package56; haplotype association testing employed haplo.stats57 with EM-based haplotype frequency estimation, adjusting for age, sex, and the first five PCs. GCTA-COJO (conditional and joint) analyses at SNCA We applied GCTA-COJO16 to perform approximate conditional (cojo-cond) and joint (cojo-joint) analyses at the SNCA locus using GWAS summary statistics and an ancestry-matched LD reference panel derived from our study genotypes after QC. Functional annotation and colocalization analysis Preliminary functional annotation is performed using SNPnexus58 (https://snpnexus.org/) and GTEx portal59 (https://gtexportal.org/home/). To further explore potential regulatory mechanisms underlying associated loci, we performed colocalization analyses using publicly available eQTL summary statistics. We leveraged eQTLGen for whole-blood eQTLs44 and brain eQTL resources—metaBrain (East Asian subset)45—for the genes of interest. Colocalization was conducted in R using the coloc package60, applying an approximate Bayes factor framework to estimate the posterior probabilities for five competing hypotheses (H0–H4), with primary interest in evidence supporting a shared causal variant between the GWAS and eQTL signals (H4). Functional annotation was performed using SNPnexus58 and GTEx Analysis Release V1059. Histone modification and broader regulatory annotation were examined using SCREEN (https://screen.wenglab.org/), which provides access to the ENCODE registry of candidate cis-regulatory elements (cCREs).61 Cross-ancestry replication and effect-size correlation To assess cross-population consistency, we extracted significant variants from the European meta-analysis (Nalls et al., 2019) and the pan-Asian GWAS (Foo et al., 2020), harmonized alleles to GRCh38, and computed Pearson correlations of log-odds ratios between this study and reference datasets. Polygenic risk score construction and site-separated evaluation We evaluated PRS models that progressively incorporated European-, East Asian-, and Taiwan-relevant variants (Fig. 6a). PRS-EUR used the published variant list and effect sizes from Nalls et al.2, after standard harmonization to our imputed dataset. PRS-EUR-Asia further added 11 independent East Asian lead variants from Foo et al.3 For these two models, published effect sizes were used directly without re-estimating SNP weights. To limit overfitting when adding Taiwan-specific variants, PRS performance was evaluated using a recruitment-site holdout design, with CGMH as the held-out test set in the primary analysis. All model fitting and score scaling were performed in the training sites only, and performance was assessed exclusively in the held-out site. We evaluated four models: PRS-EUR, PRS-EUR-Asia, PRS-TW (91 SNVs), which added LRRK2 p.G2385R, LRRK2 p.R1628P, VPS13C rs12900645, and SCARB2/FAM47E rs76591264, and PRS-TW-16 SNVs, a parsimonious model including the 11 East Asian variants, 4 Taiwan variants, and GCH1 (rs11158026). For the Taiwan-focused models, SNP weights were re-estimated in the training set by logistic regression and applied to the held-out site for out-of-sample prediction. Supplementary information Supplementary information Supplementary table7 Acknowledgements This work was conducted as part of the East Asian Parkinson Disease Genomics Consortium (EAPDGC). The EAPDGC is supported by the Michael J. Fox Foundation for Parkinson’s Research through the Genetic Diversity in Parkinson’s Disease initiative (Grant No. 17474). This project was supported by the Global Parkinson’s Genetics Program (GP2; https://gp2.org). GP2 is funded by the Aligning Science Across Parkinson’s (ASAP) initiative and implemented by The Michael J. Fox Foundation for Parkinson’s Research (MJFF). For a complete list of GP2 members see 10.5281/zenodo.7904831. For the purpose of Open Access, the author has applied a CC-BY public copyright license to any Author Accepted Manuscript version arising from this submission, the Global Parkinson’s Genetics Program (GP2), and Aligning Science Across Parkinson’s (ASAP). We also thank the Taiwan Biobank for providing control samples and support with genotype imputation, and Prof. Andrew B. Singleton and Prof. Matthew J. Farrer for their constructive comments on the analyses and manuscript. Author contributions Y.T.C. collected clinical data, performed statistical and bioinformatic analyses, and drafted the manuscript. K.Y.M. and R.M.W. conceived and designed the study. Y.A.S. performed genotyping. C.H.L., C.H.T., Y.R.W., C.T.H., and Y.W.C. recruited participants and acquired clinical data. M.H.T. supervised genotyping and data analysis. J.H. provided resources and critical intellectual input. All authors reviewed and approved the final manuscript. Data availability Summary statistics generated in this study will be made available through the GP2 platform. Access to the individual-level data used in the preparation of this article is coordinated by the corresponding authors and governed by the GP2 Tier 2 data access policy, and may be granted upon approval of the request and execution of a Data Use Agreement by the applicant’s institution. GP2 data are available via the AMP PD platform (https://amp-pd.org). Summary statistics from Nalls et al. were retrieved from the GWAS Catalog (https://www.ebi.ac.uk/gwas/), and summary statistics from Foo et al. were obtained from the supplementary materials of the original publication in JAMA Neurology.3 Publicly available eQTL summary statistics were used for colocalization analyses. Brain eQTL data were obtained from the metaBrain resource described by de Klein et al45, and whole-blood cis-eQTL summary statistics were obtained from the eQTLGen consortium website44 (cis-eQTLs; full cis-eQTL summary statistics). Code availability All code generated for this article, and the identifiers for all software programs and packages used, are available on GitHub https://github.com/GP2code/Taiwan-GWAS-and-PRS and were given a persistent identifier via Zenodo DOI 10.5281/zenodo.18232864. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Lists of authors and their affiliations appear at the end of the paper. Contributor Information Kin-Ying Mok, Email: mok@ust.hk. Ruey-Meei Wu, Email: robinwu@ntu.edu.tw. East Asian Parkinson Disease Genomics Consortium: Yung-Tsai Chu, Yu-An Su, Chin-Hsien Lin, Yih-Ru Wu, Mong-Hsun Tsai, Ruey-Meei Wu, Ji-Feng Guo, Bei-Sha Tang, Huifeng Shang, Chunyu Li, Huong Thi Thanh Nguyen, Thi Thanh Huyen Vu, Anh Tuan Nguyen, Nancy Y. Ip, Amy K. Y. Fu, Fanny C. F. Ip, Xiaopu Zhou, Han Cao, Kin Ying Mok, John Hardy, Germaine Hiu Fai Chan, Nelson Yuk-Fai Cheung, Wing-Chi Fong, Phillip Chan, Norlinah Mohamed Ibrahim, Shahrul Azmin, Nor Azian Abdul Murad, Siti Aishah Sulaiman, Siti Hajar Md Desa, Wael Mohamed, Sulev Koks, Abigail Pfaff, Lewis Singleton, Yun Joong Kim, Seok Jong Chung, Seong-Min Choi, Yu Chen, Yuewen Chen, Tao Ye, Hsiu-Chuan Wu, Tomotaka Shiraishi, Hidetomo Murakami, Renpei Sengoku, Nobutaka Hattori, Manabu Funayama, Tatsushi Toda, Wataru Satake, Kotaro Ogaki, Kenya Nishioka, Andrew B. Singleton, Mike A. Nalls, and Cornelis Blauwendraat Global Parkinson’s Genetics Program: Yung-Tsai Chu, Chin-Hsien Lin, Yih-Ru Wu, Ruey-Meei Wu, Xiaopu Zhou, Nelson Yuk-Fai Cheung, Norlinah Mohamed Ibrahim, Shahrul Azmin, Nor Azian Abdul Murad, Yun Joong Kim, Hsiu-Chuan Wu, Tomotaka Shiraishi, Nobutaka Hattori, Pin-Jui Kung, Serena Wu, Kai Shi Lim, Yi Wen Tay, Azlina Ahmad-Annuar, Tzi Shin Toh, Shen-Yang Lim, Ai Huey Tan, Jia Lun Lim, Lei-Cheng Lit, Anis Nadhirah Khairul Anuar, Hans Xing Ding, Elaine Guo Yan Chew, Jia Nee Foo, Eng-King Tan, Manabu Funayama, Hiroyo Yoshino, Rauan Kaiyrzhanov, Altynay Karimova, Gulnaz Kaishibayeva, Cholpon Shambetova, Mariam Isayan, Samson Khachatryan, Zaruhi Tavadyan, Han-Joon Kim, Jee-Young Lee, Sun Ju Chung, Beomseok Jeon, Beisha Tang, Huifang Shang, Jifeng Guo, Piu Chan, Wei Luo, Germaine Hiu-Fai Chan, Nancy Ip, Phillip Chan, Asha Kishore, K. P. Divya, Pramod Pal, Prashanth Lingappa Kukkle, Roopa Rajan, Rupam Borgohain, Mehri Salari, Bayasgalan Tserensodnom, Rajeev Ojha, Maria Leila Doquenia, Raymond Rosales, Duan Nguyen, Toan Nguyen, Maria Teresa Periñan, Paul Suhwan Lee, Mathew J. Koretsky, Sara Bandres-Ciga, Fulya Akçimen, Henry Houlden, Laurel Screven, Mie Rizig, Njideka Okubadejo, Emilia M. Gatto, Marcelo Kauffman, Claire E. Shepherd, Julie Hunter, Kishore Kumar, Melina Ellis, Miguel E. Rentería, Sulev Koks, Alexander Zimprich, Carlos Rieder, Vitor Tumas, Sarah Camargos, Edward A. Fon, Ted Fon, Oury Monchi, Benjamin Pizarro Galleguillos, Patricio Olguin, Marcelo Miranda, Maria Leonor Bustamante, Gonzalo Arboleda, Jorge Orozco, Marlene Jimenez del Rio, Alvaro Hernandez, Mohamed Salama, Walaa A. Kamel, Yared Z. Zewde, Alexis Brice, Jean-Christophe Corvol, Ana Westenberger, Christine Klein, Eva-Juliane Vollstedt, Harutyun Madoev, Joanne Trinh, Johanna Junker, Katja Lohmann, Lara M. Lange, Anastasia Illarionova, Brit Mollenhauer, Franziska Hopfner, Günter Höglinger, Manu Sharma, Thomas Gasser, Zih-Hua Fang, Sergio Groppa, Albert Akpalu, Efthymios Dardiotis, Georgia Xiromerisiou, Georgios Hadjigeorgiou, Ioannis Dagklis, Ioannis Tarnanas, Leonidas Stefanis, Maria Stamelou, Alex Medina, Andrea Quattrone, Enza Maria Valente, Micol Avenali, Lucilla Parnetti, Tommaso Schirinzi, Rejko Krüger, Wael Mohamed, Daniel Martinez-Ramirez, Mayela Rodriguez-Violante, Paula Reayes-Perez, Tim J. Anderson, Toni L. Pitcher, Oluwadamilola Ojo, Jan O. Aasly, Lasse Pihlstrøm, Manuela Tan, Shoaib Ur-Rehman, Mario Cornejo-Olivas, Angel Vinuela, Elena Iakovenko, Bashayer Al Mubarak, Muhammad Umair, Ferzana Amod, Jonathan Carr, Soraya Bardien, Esther Cubo, Ignacio Alvarez, Janet Hoenicka, Katrin Beyer, Pau Pastor, Sarah El-Sadig, Christiane Zweier, Paul Krack, Rim Amouri, Samia Ben Sassi, A. Nazl Başak, Özgür Öztop Çakmak, Sibel Ertan, Gencer Genc, Alejandro Martínez-Carrasco, Anette Schrag, Anthony Schapira, Eleanor J. Stafford, Huw Morris, John Hardy, Nicholas Wood, Olaitan Okunoye, Rimona Weil, Simona Jasaitye, Vida Obese, Camille Carroll, Claire Bale, Donald Grosset, Kin Y. Mok, Nigel Williams, Patrick A. Lewis, Seth Love, Simon Stott, Alberto Espay, Luca Marsili, Alyssa O’Grady, Bernadette Siddiqi, Bradford Casey, Brian Fiske, Charisse Comart, Justin C. Solle, Kaileigh Murphy, Maggie Kuhl, Naomi Louie, Sohini Chowdhury, Todd Sherer, Andrew K. Sobering, Cabell Jonas, Carlos Cruchaga, Caroline B. Pantazis, Claire Wegel, Deborah Hall, Ejaz Shamim, Jared Williamson, Ekemini Riley, Sonya Dumanis, Geidy E. Serrano, Thomas Beach, Honglei Chen, Ignacio Juan Keller Sarmiento, Niccolò E. Mencacci, Steven Lubbe, Joseph Jankovic, Miguel Inca-Martinez, Joshua Shulman, Karen Nuytemans, Karl Kieburtz, Katerina Markopoulou, Kenneth Marek, Lana M. Chahine, Lauren Ruffrage, Marissa Dean, Lisa Shulman, Roger Albin, Roy Alcalay, Ruth Walker, Tao Xie, Tatiana Foroud, and Masharip Atadzhanov Supplementary information The online version contains supplementary material available at 10.1038/s41531-026-01349-0. References Associated Data Supplementary Materials Supplementary information Supplementary table7 Data Availability Statement Summary statistics generated in this study will be made available through the GP2 platform. Access to the individual-level data used in the preparation of this article is coordinated by the corresponding authors and governed by the GP2 Tier 2 data access policy, and may be granted upon approval of the request and execution of a Data Use Agreement by the applicant’s institution. GP2 data are available via the AMP PD platform (https://amp-pd.org). Summary statistics from Nalls et al. were retrieved from the GWAS Catalog (https://www.ebi.ac.uk/gwas/), and summary statistics from Foo et al. were obtained from the supplementary materials of the original publication in JAMA Neurology.3 Publicly available eQTL summary statistics were used for colocalization analyses. Brain eQTL data were obtained from the metaBrain resource described by de Klein et al45, and whole-blood cis-eQTL summary statistics were obtained from the eQTLGen consortium website44 (cis-eQTLs; full cis-eQTL summary statistics). All code generated for this article, and the identifiers for all software programs and packages used, are available on GitHub https://github.com/GP2code/Taiwan-GWAS-and-PRS and were given a persistent identifier via Zenodo DOI 10.5281/zenodo.18232864.