Variation in PARK10 is not associated with risk and age at onset of Parkinson’s disease in large clinical cohorts

Abstract

A recent study in autopsy-confirmed Parkinson’s disease (PD) patients and controls revived the debate about the role of PARK10 in this disorder. In an attempt to replicate these results and further understand the role of this locus in the risk and age at onset of PD, we decided to explore NeuroX genotyping and whole exome sequencing data from 2 large independent cohorts of clinical patients and controls from the International Parkinson’s Disease Genomic Consortium. A series of single-variant and gene-based aggregation (sequence kernel association test and combined multivariate and collapsing test) statistical tests suggested that common and rare genetic variation in this locus do not influence the risk or age at onset of clinical PD.

1. Introduction

Whether genetic variation across PARK10 locus is associated with Parkinson’s disease (PD) has been an area of much discussion. This locus in chromosome 1p32.3 was first associated with PD in 2003 in a genome-wide scan in a set of Icelandic families (Hicks et al., 2002). Subsequently, this locus was shown to influence age at onset (AAO) of PD in a genomic screen of several unrelated families (Oliveira et al., 2005). After its identification, there has been a lot of controversy about the real implication of PARK10 with either risk or AAO of PD, with some studies identifying association and others being largely negative (Farrer et al., 2006; Haugarvoll et al., 2009; Li et al., 2007; Maraganore et al., 2005; Wan et al., 2014). One of the studies suggesting a role of this locus in PD was the very first PD genome-wide association study (GWAS), identifying 2 associated single-nucleotide polymorphisms (SNPs) (rs682705 and rs7520966) after a 2-tier analysis (Maraganore et al., 2005). The following 9 years have witnessed numerous PD GWASs including 3 large meta-analyses derived from large international collaborations (International Parkinson Disease Genomics Consortium et al., 2011; International Parkinson’s Disease Genomics Consortium and Wellcome Trust Case Control Consortium, 2011; Nalls et al., 2014). These studies have identified up to 28 independent loci that modulate the risk of PD, demonstrating an unequivocal role for common genetic variation in the etiology of this disorder, and pointed to SNCA, MAPT, GBA-SYT11, HLA-DQB1, and GAK-DGKQ as major risk loci. However, none of these studies have shown significant association results around PARK10. Although these studies have been successful in identifying common variants associated with PD, they only explain a fraction of the expected heritability of PD (Keller et al., 2012). Thus, common variants of smaller effect and rare variants of moderate effects remain to be discovered. In a recent publication by Beecham et al. (2015), the authors, hypothesized that one of the reasons for this missing heritability may be the uncontrolled degree of neuropathologic (and therefore etiological) heterogeneity in clinical series of PD cases. To reduce this heterogeneity, they performed a GWAS in which only PD cases with autopsy-confirmed Lewy body (LB) pathology and controls without neuropathology were included. Although they used a relatively small number of samples, they found evidence suggesting that common variation in PARK10 is associated with LB PD (rs10788972, p = 6.2 × 10⁻⁸) with an odds ratio for the protective allele comparable with that found for SNCA and MAPT in other studies. They demonstrated that rs10788972 is in strong linkage disequilibrium with the SNP defining the PARK10 haplotype previously shown to be significantly associated with AAO in PD. The critical region containing the PARK10 locus was significantly reduced from 10.6 Mb to 100 kb and contains only 4 known genes: TCEANC2, TMEM59, miR-4781, and LDLRAD1 (Beecham et al., 2015).

To further understand the role of common and rare variation around this locus in the risk and AAO in PD, we decided to explore the large genetic repository of the International Parkinson’s Disease Genomic Consortium (IPDGC), consisting of NeuroX genotyping data from 7155 PD cases and 6480 controls, and an independent dataset of whole exome sequencing (WES) data of 1189 cases and 469 controls.

2. Materials and methods

2.1. NeuroX genotyping

NeuroX is a semi-custom genotyping array containing ~240,000 exonic variants available on the Illumina Infinium HumanExome BeadChip and an additional ~ 24,000 variants proven or hypothesized to be relevant in neurodegenerative diseases (Nalls et al., 2015). Of these neurodegenerative-related variants, more than 9000 are dedicated to PD, including tagging SNPs, proxies, and technical replicates of 28 genome-wide significant loci from the discovery phase of a recent large meta-analysis (Nalls et al., 2014). Genotyping data from a large set of samples partially overlapping with those used in the replication stage of the mentioned meta-analysis were used for this study. After extensive quality control (described elsewhere Nalls et al., 2014), a total of 45 variants in 7155 PD cases and 6480 controls were called within 100 kb upstream and downstream of the critical region defined by Beecham et al. (2015). These variants were extracted with VCFtools, version 0.1.13 (https://vcftools.github.io/index.html), and assigned gene-based and impact score annotations with KGGSeq (http://statgenpro.psychiatry.hku.hk/limx/kggseq/) using refseq hg19 and dbNSFP (https://sites.google.com/site/jpopgen/dbNSFP) databases. Based on these annotations, variants were classified as functional-1 (frameshift, nonframeshift, startloss, stoploss, stopgain, splicing, missense, variants within the 5′ and 3′ untranslated regions, and those present 100 base pairs upstream and downstream of a transcription start or end site, respectively), functional-2 (frameshift, nonframeshift, startloss, stoploss, stopgain, splicing, and missense), and/or functional-3 (variants predicted to be Mendelian disease causal by a logistic regression model as described elsewhere, Li et al., 2013).

2.2. Whole extime sequencing

Sample libraries were prepared with Roche Nimblegen or Illumina exome capture kits and subjected to 100-base pair paired-end sequencing on the Illumina HiSeq2000. Sequence reads were aligned to the reference genome (hg19) using the Burrows-Wheeler Aligner (BWA)-MEM algorithm of the BWA software package 0.7.9a (http://bio-bwa.sourceforge.net). Picard tools, version 1.129 (http://broadinstitute.github.io/picard/), was used to create bam files and to sort and index the sequence reads. Single-nucleotide variants and insertion/deletions were called and recalibrated using the Genome Analysis Toolkit, version 3.3-0 (https://www.broadinstitute.org/gatk/), following the recommended workflow for variant analysis. After extensive quality control (see Supplementary Materials) a total of 101 variants in 1189 PD cases and 469 controls were called within 100 kb upstream and downstream of the critical region defined by Beecham et al. (2015). These variants were extracted and annotated with VCFtools and KGGSeq, respectively, and grouped into 3 different functionality groups as explained earlier.

2.3. Power calculations for single-variant tests

Power calculations for single-variant tests were computed with Quanto, version 1.2.4 (http://biostats.usc.edu/Quanto.html). A gene-only effect, an additive mode of inheritance, and a population risk of 1.8% (de Rijk et al., 2000) were assumed. A significance level of 0.0011 (0.05/45) and 0.000495 (0.05/101) were used for the NeuroX and WES datasets, respectively. The size of effect reported by Beecham et al. (2015) was calculated for pathologically confirmed PD cases and controls. Because we cannot rule out the possibility that a fraction of the samples in our cohort may not fulfill the inclusion criteria used by the authors (affecting the allele frequency and size of effect of the associated variant), the mentioned power calculations were performed at different disease allele frequencies (0.1–1.0) and sizes of effect. Thus, odds ratios ranging from 0.1 to 2.0 were used for the dichotomous model, whereas main effects ranging from 0 to 12 were used for the continuous model.

2.4. Power calculations for burden tests

Power for gene-base aggregation tests was estimated using sequence kernel association test (SKAT), version 1.0.9, package for R (http://www.r-project.org/). The total length of the coding regions of the transcripts in this locus was used as subregion length. Three different minor allele frequency (MAF) cutoffs for the causal variant were used (0.05, 0.01, and 0.001). This analysis was repeated considering 2 and 5 causal variants among the total number of rare variants. Power was computed for a total of 1000 causal variant/subregion sets from the SKAT haplotype dataset, and the average power was obtained by taking the mean of the computed powers. The level of significance used was 0.00625 (0.05 divided by the number of transcripts in this region). These analyses were applied to the complete sets of NeuroX and WES variants and to those variants contained in the 3 functionality groups defined earlier.

2.5. Logistic regression models

Each of the variants identified was tested for association with PD with a logistic regression analysis in PLINK 1.9 (https://www.cog-genomics.org/plink2). For variants genotyped with the NeuroX array, gender and the first 10 components after multidimensional scaling (MDS) analysis were used as covariates. For variants identified with WES, gender, the kit used for the capture experiments, the percentage of base pairs with at least 10× and 30× coverage, and the first 10 MDS-analysis components were used.

2.6. Linear regression models

To test the influence of the identified variants in the AAO of PD, linear regression analyses were performed with PLINK 1.9, using AAO of PD as the dependent variable. Data from the control individuals were not included in these analyses. For variants genotyped with the NeuroX array, gender and the first 10 MDS-analysis components were used as covariates. For variants identified with WES, gender, the kit used for the capture experiments, the percentage of base pairs with at least 10× and 30× coverage, and the first 10 MDS-analysis components were used.

2.7. SKAT analysis

A total of 8 complete transcripts are present within the region understudy -chr1: 54,372,370-54,679,030 (hg19)-, namely HSPB11, LRRC42, LDLRAD1, TMEM59, TCEANC2, miR-4781, CDCP2, and CYB5RL. To test whether the joint burden of common and rare variants within each of these transcripts is associated with a PD or AAO in this disease, SKATs at different common-rare frequency cutoffs (0.05, 0.01, and 0.001) were applied. These analyses were applied to the entire sets of NeuroX and WES variants and variants within the 3 functional groups defined earlier. For variants genotyped with the NeuroX array, gender and the first 10 MDS-analysis components were used as covariates. For variants identified with WES, gender, the kit used for the capture experiments, the percentage of base pairs with at least 10× and 30× coverage, and the first 10 MDS-analysis components were used. In all instances, SKAT version 1.0.9 package for R was used.

2.8. Combined multivariate and collapsing test

We also performed a combined multivariate and collapsing (CMC) test that assumes all variants to have the same direction of effect and is a powerful approach to assess the main effects of susceptibility genes in complex traits. This test was applied to the entire sets of NeuroX and WES variants and variants within the 3 functional groups described earlier. For variants genotyped with the NeuroX array, gender and the first 10 MDS-analysis components were used as covariates. For variants identified with WES, gender, the kit used for the capture experiments, the percentage of base pairs with at least 10× and 30× coverage, and the first 10 MDS-analysis components were used. Association with PD and AAO of PD was investigated in all instances. These test were performed with Rvtests (http://zhanxw.github.io/rvtests/).

3. Results

To further understand the role of PARK10 in the risk and AAO of clinical PD, we decided to mine NeuroX genotyping and WES data from the IPDGC. Thus, variation around the critical region defined by Beecham et al. (2015) was extracted from these 2 independent datasets. A total of 45 variants genotyped in 7155 PD cases and 6480 controls were selected from the NeuroX dataset, whereas 101 variants sequenced in 1189 PD cases and 469 controls were selected from the WES dataset (Table 1). A detailed list of the variants identified is shown in Supplementary Tables 1 and 2.

Table 1.

Characteristics of NeuroX and WES datasets

Dataset	Variants assessed	Cases			Controls
		Sample size	AAO, mean (SD)	M/F ratio	Sample size	AAE, mean (SD)	M/F ratio
NeuroX	45	7155	61.06 (12.60)	1.77	5480	63.76 (15.54)	1.26
WES	101	1189	42.21 (11.32)	1.5	469	46.14 (26.49)	1.83

Key: AAE, age at examination; AAO, age at onset; F, female; M, male; SD, standard deviation; WES, whole exome sequencing.

A series of power calculations were performed to estimate the capacity of these datasets to detect association of PARK10 with a risk or AAO of PD. These tests showed that the NeuroX data have 99.9% power to detect single variation associated with PD at the MAF and size of effect values attributed to PARK10 by Beecham et al. (2015) (Supplementary Fig. 1). This dataset has also excellent power (91.9%–99.9% depending on the main effect of the associated variant) to detect single variation associated with AAO of PD at this same MAF cutoff (Supplementary Fig. 2). The power of this dataset to detect a significant burden of variants in transcripts in this locus in PD cases versus controls is also excellent depending on the causal MAF cutoff used, the number of causal variants considered to be present in this locus, and the level of functionality used for the calculation (Supplementary Table 3). Similarly, the power to detect association between burden of variants in these transcripts and AAO of PD can be excellent depending on the variables mentioned earlier (Supplementary Table 4).

The WES dataset has in general less power than the NeuroX dataset as it comprised a smaller number of both PD cases and controls. However, the capability of this dataset can also be excellent, depending on the variables used for the analysis. Thus, this dataset has 91.3% power to detect single variants associated with PD at the MAF and size of effect values assigned to PARK10, and 10.1%–99.9% power to detect association with AAO, depending on the main effect of the associated variant (Supplementary Figs. 1 and 2). The power of this dataset to detect a significant burden of variants in transcripts in this region is significantly lower, with a maximum of 83.4% if the causal MAF cutoff is 0.05, the number of causal variants in the region is 5, and only variants within the third functionality group (see Section 2) are considered (Supplementary Table 3). The power to detect an association between burden of variants in these transcripts and AAO of PD has a maximum of 95.1% if the same assumptions are made (Supplementary Table 4).

To check the possibility that single variation is associated with a risk or AAO of PD in our dataset, multicovariate regression models were applied to each variant as described in Section 2. None of these tests derived significant results after multiple-test correction. Furthermore, those variants nominally associated with a risk or AAO of PD did not replicate across the 2 datasets understudy. These results altogether indicate that single variation is not associated with PD or AAO in PD in either of these cohorts (Supplementary Tables 5 and 6).

Because most of the variants identified are rare, achieving adequate statistical power requires combination and evaluation of groups of variants likely to have similar function. Thus, we decided to test whether the burden of common and rare variants within transcripts in this region is associated with PD or AAO of PD. For this purpose, we chose to use SKAT at different common-rare frequency cutoffs (0.05, 0.01, and 0.001). These analyses were performed for NeuroX and WES variants independently and drew no significant results after correcting for the number of tests (number of genes present in this region) performed (Fig. 1 and Supplementary Tables 7-10). Furthermore, hypothesizing that the mentioned burden might only be true for certain groups of functional variants, both datasets were subsetted according to different definitions of functionality (see Section 2). Following these definitions, NeuroX variants were subsetted in 3 groups (functional 1–3) containing 42, 42, and 15 variants, respectively. WES functionality groups contained 69, 57, and 17 variants, respectively. Again, SKATs at different common-rare frequency cutoffs derived no significant association results after multiple-test corrections. To note, nominally associated results were not consistent across the 2 datasets used herein, further supporting that burden of variants in transcripts contained in this region is not associated with either risk or AAO of clinical PD (Fig. 1 and Supplementary Tables 7-10).

Fig. 1.

Sequence kernel association tests were used to test whether the burden of common and rare variants within transcripts in this region is associated with a risk (A and C) or age at onset (AAO) (B and D) of Parkinson’s disease (PD). Panels (A) and (B) show results obtained for those variants genotyped with NeuroX array, whereas (C) and (D) the results obtained for whole exome sequencing (WES)-identified variants. These tests were performed at different common-rare frequency cutoffs (0.05, 0.01, and 0.001) and under different functionality definitions (see Section 2). −Log p values derived from these tests are represented by color-coded bars in each of the panels. The threshold for significance after correcting for the number of transcripts present in this region is shown as a dotted red line.

The aforementioned SKAT is more powerful for causal variants with different directions of effect. In the case of PARK10, it could be assumed from previous studies that all variants tend to be causal. Therefore, a test that assumes that all variants have the same direction of effect could be a more powerful approach to assess the role of PARK10 in PD. For this reason, a series of CMC tests considering the 3 different definitions of functionality previously mentioned were applied to the NeuroX and WES variants independently. None of these tests yielded significant association results after correcting for the number of test performed (Supplementary Tables 11 and 12).

4. Discussion

A recent publication by Beecham et al. (2015) reignited the debate about the role of PARK10 in PD. In this study, the authors identified evidence that common variation in this locus is associated with autopsy-confirmed PD, with a size of effect similar to that identified in SNCA or MAPT in other studies. These results are in contrast with GWAS efforts failing to identify significant association at this locus (International Parkinson Disease Genomics Consortium et al., 2011; International Parkinson’s Disease Genomics Consortium and Wellcome Trust Case Control Consortium, 2011; Nalls et al., 2014). One of such efforts is a recent meta-analysis in an unprecedented number of cases and controls that identified genome-wide significant association at 28 independent loci and failed to find significant association around PARK10 (Nalls et al., 2014). This study included only clinical cases and both genetic heterogeneity and the number of tests performed could be masking the effect of loci associated with specific subtypes of PD. However, it is surprising that a reduction of the number of tests by post hoc analysis does not yield significant results at this locus (Supplementary Fig. 3 and Supplementary Table 13), especially after considering its attributed size of effect and the vast number cases and controls used in this study. Given this controversy, we decided to mine NeuroX genotyping and WES data from 2 large independent cohorts of PD patients and controls from the IPDGC.

A series of logistic regression and burden (SKAT and CMC) tests showed that common and rare variation in this locus do not modulate the risk of PD in either of these datasets. Furthermore, because this locus has previously been associated with the AAO in PD, linear regression models and the mentioned burden tests were performed using AAO data as a continuous trait. Results derived from these analyses suggested that common and rare variation do not influence the AAO in the analyzed PD cases.

In summary, results presented herein, along with previously published data, suggest that PARK10 is not associated with risk or AAO of disease in large clinical series of PD. It could be argued that this locus is only associated with a specific subgroup of PD and that its size of effect might be diluted in the large clinical series presented herein. However, power calculations performed at different size of effect and MAF cutoffs suggest that even considering this scenario, the datasets presented in this study have sufficient power to detect association at this locus. For this reason, because 85% diagnosis accuracy is expected in clinical series of PD (Adler et al., 2014) and because most of these cases would meet the inclusion criteria used by Beecham et al. (2015), it is tempting to conclude that variation in this locus is not associated with PD as a whole. However, we have to be cautious and cannot rule out the possibility that PARK10 is associated with LB-confirmed PD. To confirm or reject this possibility, further association studies including autopsy-confirmed PD cases and controls are needed.