Subjects were enrolled in the large cohort of Parkinson's Disease and Movement Disorders Multicenter Database and Collaborative Network in China (PD-MDCNC, http://www.pd-mdcnc.com/), as described in a previous study (12). Patients with PD in this cohort were diagnosed by experienced neurologists according to the UK Parkinson's Disease Scoeity (PDS) Brain Bank Criteria (13) or Movement Disorder Society Clinical Diagnostic Criteria (14). Meanwhile, neurological disease-free controls were recruited, with matched ethnicity as a reference. As indicated in our earlier study, we excluded individuals with pathogenic/likely pathogenic mutations of 23 PD pathogenic genes from the EOPD and FPD cohorts. These patients were divided into two separate cohorts, which were named according to the sequencing method as follows: (1) the whole-exome sequencing (WES) cohort, containing 1,917 familial or sporadic early-onset PD from Mainland China (mean age, 52.22 ± 9.03 years; women, 45.4%) and 1,652 race-matched healthy controls (mean age, 62.03 ± 12.59 years; women, 51.9%), and (2) the whole-genome sequencing (WGS) cohort, containing 1,962 sporadic late-onset PD from Mainland China (mean age, 66.76 ± 7.078 years; women, 49.8%) and 1,652 race-matched healthy controls (mean age, 62.32 ± 7.109 years; women, 52.1%). The WES cohort consisted of 477 familial PD probands (327 AD probands and 150 AR probands) and 1,440 patients with early-onset PD (age of onset was not more than 50 years). The WGS cohort contained patients with late-onset sporadic PD with mean age onset of 61.88 ± 6.927 years. Patients in both cohorts carrying pathogenic mutations in high-confidence PD-causing genes were excluded from the analysis. Followed by informed consent, the basic demographic data, peripheral blooding samples, and clinical features of all participants were collected. We collected total genomic DNA from peripheral blooding samples using standard procedures. The respective ethics committees of Xiangya Hospital of Central South University approved the abovementioned protocol.

As shown in our previous study (15), the WES cohort used SureSelect Human All Exon Kit V6 (Agilent) to capture the whole-exome DNA, prepared the sample library, and then used Illumina HiSeq 10× for pair-end 2 × 150 bp sequencing. The average sequencing depth was 123×, achieving a coverage of at least 10× for 99.32% of the target region. WGS was performed using the Illumina Nova Sequencing platform in a pair-end 2 × 150 bp mode, and the average depth of coverage was about 12×. Sequencing data of both groups were processed and analyzed with the BWA-GATK-ANNOVAR pipeline (16, 17). Quality control was conducted as described in our previous study (15). Samples would be removed if they had sex discrepancies, abnormal heterozygosity (>3 SD), pathogenic or likely pathogenic variants of PD-related genes, or unusual relatedness (descent > 0.15). In addition, we performed the principal component analysis (PCA) using PLINK v1.90 to assess potential population structure stratification. In subsequent analyses for the WGS cohort, gender, age, and the first five principal components of population stratification were used as covariates, whereas for the WES cohort (the control group included elderly people without neurological diseases), sex and the first five principal components for population stratification were used (18).

In both two cohorts, we targeted the variants in the coding region of the DNM1L gene (NM_001278466). The variants with a missing rate of >5% and deviations from Hardy–Weinberg equilibrium in controls (p < 0.05) were removed using PLINK v1.90. We annotated the gene regions (hg19 RefSeq), amino acid changes, and allele frequency of each variant in the Genomic Aggregation Database (gnomAD) and Exon Aggregation Consortium (ExAC) by ANNOVAR. Next, the functional impact of each nonsynonymous variant was predicted by ReVe (threshold, 0.7).

We categorized the variants according to the minor allele frequency (MAF) as common variants (MAF > 0.01) and rare variants (MAF < 0.01). Furthermore, we re-extracted the rare variants with MAF below 0.001 and performed gene analysis of rare variants with MAF below 0.01 and 0.001, respectively. According to predicted functions, all the rare nonsynonymous variants were classified into three variant groups: missense, potentially damaging missense (Dmis, ReVe score > 0.7), and loss of function variants (LoF stop gain/loss, frameshift, and splice site), and the sum of Dmis and LoF. After adjusting the abovementioned covariates in two cohorts, sequence kernel association test-optimal (SKAT-O) was applied to each cohort to assess the combined effect of rare variants and each variants group. Fisher's exact test was also done for common variants to validate the significant relationship between the common variants of DNM1L and PD. Moreover, logistic regression analysis based on an allele model was also performed by PLINK v1.90, and a p-value of < 0.05 was considered suggestive significant.

To confirm the involvement of DNM1L variants in PD susceptibility, a meta-analysis combining public studies and our case–control study was conducted. In addition to our data, summary data from the PD variant browser were included in the meta-analysis, which comprised four studies: PD Genome Project, International Parkinson's Disease Genomic Consortium (IPDGC) Exomes, IPDGC Resequencing Project, and UK Biobank (19). Unlike our cohorts, these cohorts mainly contain European populations, which can increase the power to detect the association between DNM1L variants and PD. As no rare variants of DNM1L found in our cohort were seen in included public data, we only validated the significant association between significant common variants of DNM1L and PD in meta-analysis. We used the Hardy-Weinberg equilibrium model to estimate the SNPs in all cohorts and then excluded the variants with deviations in controls (p < 0.05). To assess the strength of the association between DNM1L variants and PD risk, a pooled OR and 95% confidence intervals (CIs) were calculated under five different models (allele, dominant, recessive, heterozygote, and homozygote model). The Cochrane Q-test and I² statistics were used to assess study heterogeneity and a significant Q-test (p < 0.1 or I² > 50%) indicated heterogeneity. Fixed- or random-effects models were selected based on the presence or absence of heterogeneity. A Z-test determined the significance of the pooled ORs. We used the FDR method to correct p-values for multiple comparisons for the variant association analysis. A p-value of < 0.05 was considered statistically significant. Moreover, the p-values of Egger's and Begg's tests were calculated to estimate the publication bias. All analyses were performed using the R package “meta.”

The basic demographic characteristics of the two cohorts are summarized in Supplementary Table S1. In the WES cohort, a total of 1,917 familial PD probands or sporadic early-onset cases and 1,652 controls were included. The WGS cohort included 1,962 sporadic late-onset PD (sLOPD) patients and 1,279 healthy controls. After variant filtering and classification, we identified 224 variants in the coding region of DNM1L, including 23 rare nonsynonymous variants and 201 common variants.

As shown in Figure 1 and Table 1, the MAF of the 23 rare variants was below 0.001. Of these 23 rare variants, 21 of them were missense mutations, and 12 were predicted to be damaging. Of the 12 damaging missenses, five of them were only found in patients with PD (p.V213A, p.L116Q, p.G120S, p.P255R, and p.P411L). In addition, we found two shear mutations that may lead to loss of function (c.987+2_987+3insA and c.988-1_988insTCT). In detail, c.987+2_987+3insA was detected in one control from the WES cohort and in two patients from the WGS cohort, whereas c.988-1_988insTCT was detected in only one patient from the WGS cohort.

Table 2 shows the results of the rare variants association test of DNM1L. To increase the power to detect association, we performed the burden analyses on WES and WGS cohorts stratified by the functional effects of variants. As a result, no relationship between PD and variants was observed in any variants groups (p > 0.05, Table 2).

Among the 201 common variants, only two common variants (rs10844308 and rs143794289) in the WES cohort were significant in the Fisher's test. Also, rs10844308 was suggested to cause a 22% increase in the odds of PD (OR = 1.220, p = 0.036). On the contrary, rs143794289 was presented as a protective factor of PD since it caused a 28.2% decrease in the odds of PD (OR = 0.718, p = 0.025). However, the result of these two SNPs turned negative in the WGS cohort analysis. Logistic regression with PLINK also analyzed associations between common variants and PD. Age at enrollment, sex, and the first five principal components was used as adjustment covariates. Nevertheless, logistic regression did not demonstrate any suggestive result between DNM1L and PD (p > 0.05, Supplementary Tables S2, S3).

As mentioned earlier, we found two significant common variants in the WES cohort—rs10844308 and rs143794289. These two SNPs were further included in the meta-analysis. For all included studies, genotypes of these two variants in controls were consistent with Hardy–Weinberg equilibrium (p > 0.05). With regard to rs10844308, data were obtained from six cohorts (WES, WGS, and four public cohorts), including 28,985 patients and 69,206 controls. In both the Chinese population and European populations, no genetic models showed evidence of heterogeneity (p for heterogeneity > 0.1, I² < 50). Therefore, we applied a fixed-effects model for all the analyses of rs10844308. For the meta-analysis of all included studies, no genetic model indicated the significance of rs10844308 in PD risk under the fixed-effect model, and then, an analysis stratified by ethnicity was conducted. In the Chinese population, only the allele model (C vs. A) presented significantly increased PD risk with an OR of 1.18 (95% CI = 1.03–1.35). However, in the European population, all genetic models failed to show the significance of rs10844308 in PD risk (Table 3, Supplementary Figure).

The results of rs143794289 turned out to be more complicated but interesting. Unlike rs10844308, the homozygous mutation of rs143794289 was not found in public data, so we only applied the allele, dominant, and heterozygote genetic models to analyze the association between rs143794289 and susceptibility of PD. For rs143794289, we included four studies (the WGS cohort, the WES cohort, and two public cohorts) containing 15,072 patients and 20,541 controls. In the overall analysis, all genetic models showed no evidence of heterogeneity (p for heterogeneity > 0.1, I² = 0) and presented significantly reduced risk of PD, with ORs of 0.80 (95% CI = 0.66–0.97), 0.81 (95% CI = 0.66–0.98), and 0.80 (95% CI = 0.66–0.97) for the allele, dominant, and heterozygote genetic models, respectively. After correction using the FDR method, the results of dominant and heterozygote genetic models remained significant (p = 0.027 and 0.035, respectively). Furthermore, in the Chinese population, all three genetic models show a protective role of rs143794289 in PD, while no significant result was found in the European population (Table 3, Supplementary Figure). All p-values from Egger's and Begg's tests were >0.05, indicating no publication bias in this meta-analysis.