No Evidence for a Causal Relationship Between Cancers and Parkinson’s Disease

Abstract

Background:

Epidemiological data suggest that cancer patients have a reduced risk of subsequent Parkinson’s disease (PD) development, but the prevalence of PD in melanoma patients is often reported to be increased. Causal relationships between cancers and PD have not been fully explored.

Objective:

To study causal relationship between different cancers and PD.

Methods:

We used GWAS summary statistics of 15 different types of cancers and two-sample Mendelian randomization to study the causal relationship with PD.

Results:

There was no evidence to support a causal relationship between the studied cancers and PD. We also performed reverse analyses between PD and cancers with available full summary statistics (melanoma, breast, prostate, endometrial and keratinocyte cancers) and did not find evidence of causal relationship.

Conclusion:

We found no evidence to support a causal relationship between cancers and PD and the previously reported associations could be a result of genetic pleiotropy, shared biology or biases.

INTRODUCTION

Parkinson’s disease (PD) is a complex disorder, influenced by numerous environmental and genetic factors. Observational studies have suggested associations between PD and different types of cancers (lung, skin, pancreatic cancers and others) [1–7], such that cancer patients have lower risk of subsequent PD development [8] and overall PD is associated with a reduced risk of subsequent cancer development [1, 2], However, risk of PD is increased in melanoma patients [9] and the prevalence of melanoma and brain tumors may be increased in patients with PD [3–6]. In the absence of a causal effect, apparent associations may be explained by confounding factors (such as toxins that casually influence the risk of specific cancers and PD), shared genetic susceptibility or biological pathways, or ascertainment bias [10,11].

In Mendelian randomization (MR), similar to randomized control trials, single-nucleotide polymorphism (SNPs) are used to randomly divide participants into two groups defined by genotype, assuming that genotype distribution is a random process during meiosis, and therefore it should not be affected by confounders. MR uses SNPs associated with an exposure of interest (such as cancer susceptibility) as proxies to determine the causal association between that exposure and an outcome [12]. Summary level data from genome wide associational studies (GWASs) are used to construct instrumental variables (IVs) from GWAS significant SNPs. In the current study, we performed bi-directional MR to examine whether certain types of cancers have causal relationships with PD and vice versa.

METHODS

Mendelian randomization

For the construction of genetic instruments, we selected studies from the GWAS Catalog [13] using the R package MRInstruments [14, 15]. First, we searched for traits using keywords “cancer”, “carcinoma”, “glioma”, “lymphoma”, “leukemia”, “melanoma”. We then selected the most recent available GWAS for each cancer, with a minimum of 1000 cases and at least the same number of controls of European ancestry. Additionally, recent GWASs on melanoma [16] and combined analysis of keratinocyte cancers [17] were added as they were not available in the GWAS catalog. Fifteen studies were selected for this part of the analysis (Supplementary Table 1). UK biobank (UKB) participants were included in some of these studies (colorectal cancer, combined analysis of keratinocyte cancers, endometrial cancer, lung cancer, melanoma, uterine fibroids).

To perform MR in the reverse direction (the causal relationship between PD and different cancer types) we required full summary statistics which we obtained through GWAS Catalog or direct contact with authors. We were able to collect full summary statistics for melanoma [16], breast [18], prostate [19], endometrial [20] and keratinocyte cancers (basal cell and squamous cell carcinoma) [17].

We used GWAS summary statistics from the latest PD GWAS excluding 23andMe and UKB data, to avoid potential bias due to overlapping samples [21]. After the exclusions, a total of 15,056 PD patients and 12,637 controls were included in the summary statistics [21].

We constructed genetic instruments for cancer susceptibility and PD using SNPs with GWAS significant p-values (< 5 × 10⁻⁸) from each study. The extracted data included rs-numbers, log odds ratios, standard errors, p-values, alleles, and effect allele frequency. SNPs for each exposure were clumped using standard parameters (clumping window of 10,000 kb, r² cutoff 0.001) to discard variants in LD. Additionally, we calculated r², which reflects the proportion of variability explained by genetic variants and F-statistics to estimate the strength of IVs selected for exposures as previously described [22, 23]. We calculated estimated power to detect an equivalent effect size of OR 1.2 on PD risk utilizing an online Mendelian randomization power calculation (https://sb452.shinyapps.io/power/) [24].

MR methods implemented in the Two-sample MR R package [14, 15] were used and are described in detail elsewhere [25–27]. Firstly, we performed Steiger filtering to exclude SNPs that explain more variance in the outcome than in the exposure [15]. We then used the inverse variance weighted (IVW) method, in which we pooled estimates from individual Wald ratios for each SNP and meta-analyzed using random effects [25–27]. We applied MR Egger to detect net directional pleiotropy and provide a better estimate of the true causal effect allowing to detect possible violations of instrumental variable assumptions [27]. Additionally, we used weighted median (WM) which is a median of the weighted estimates and provides consistent effect even if 50% of IVs are invalid [28], These sensitivity analyses were performed to explore heterogeneity and horizontal pleiotropy. Heterogeneity was tested using Cochran’s Q test in the IVW and MR-Egger methods [29], For each method, we constructed funnel plots to detect pleiotropic outliers (Supplementary Figures 1–6). Additionally, we performed MR-PRESSO test to detect outlier SNPs which may be biasing estimates through horizontal pleiotropy, and then adjust for them [30].

Data availability

All code used in the current study is available at https://github.com/gan-orlab/MR_Cancers-PD

RESULTS

Mendelian randomization does not support a causal role for different cancers and PD

We selected 15 cancer GWAS studies for MR analysis (Table 1). The variance in the exposure variables explained by SNPs ranged from 0.016 to 0.059 (Table 2). All instruments had F-statistics of > 10, which is the standard cut-off applied to indicate sufficient instrument strength (Table 2; Supplementary Table 1).

Table 1.

List of all cancer GWASs selected for Mendelian randomization analysis

Trait	Study	Initial sample size		Replication sample size		Power
		Cases	Controls	Cases	Controls
Breast cancer	Michailidou et al., 2017 [18]	76,192	63,082	46,785	70,064	100.00%
Chronic lymphocytic leukemia	Law et al., 2017 [43]	4,478	13,213	1,722	4,385	80.70%
Colorectal cancer	Law et al., 2019 [44]	31,197	61,770	–	–	38%
Cutaneous squamous cell carcinoma	Chahal et al., 2016 [45]	6579	280,558	825	11,518	74.50%
Combined analysis of keratinocyte cancers	Liyanage et al., 2019 [17]	31,787	619,351	–	–	63.00%
Endometrial cancer	O’Mara et al., 2018 [20]	12,906	108,979	–	–	71.50%
Lung cancer	McKay et al., 2017 [46]	23,223	16,964	–	–	71.50%
Lymphoma	Sud et al., 2017 [47]	1,278	14,325	1,586	3,069	90.60%
Melanoma	Landi et al., 2020 [16]	36,760	375,188	–	–	68.30%
Non-glioblastoma glioma/Glioma	Melin et al., 2017 [48]	12,469	18,190	–	–	93.10%
Oral cavity and pharyngeal cancer	Lesseur et al., 2016 [49]	6,009	6,585	–	–	95.60%
Pancreatic cancer	Klein et al., 2018 [50]	9,040	12,496	2,737	4,752	82.80%
Prostate cancer	Schumacher et al., 2018 [19]	79,148	61,106	–	–	57.00%
Renal cell carcinoma	Scelo et al., 2015 [51]	10,784	20,406	3,182	6,301	71.50%
Uterine fibroids	Rafhar et al., 2018 [52]	16,595	52,3330	–	–	64.90%

Table 2.

MR analysis between exposure (cancers) and outcome (PD)

Exposure	N, SNPs included	r2	F-statistics	MR Egger			Inverse variance weighted
				b	se	p	b	se	p
Breast cancer	107	0.016	38.5	0.075	0.065	0.247	0.032	0.033	0.337
Chronic lymphocytic leukemia	7	0.035	106.11	0.047	0.640	0.944	0.099	0.077	0.197
Colorectal cancer	35	0.02	53.8	−0.002	0.273	0.994	0.042	0.057	0.460
Cutaneous squamous cell carcinoma	23	0.03	405.2	−0.097	0.077	0.223	0.051	0.048	0.288
Combined analysis of keratinocyte cancers	68	0.023	216.6	−0.018	0.053	0.732	0.017	0.031	0.586
Endometrial cancer	13	0.028	271.4	−0.106	0.252	0.681	−0.014	0.059	0.808
Lung cancer	10	0.029	120.4	0.000	0.121	0.999	0.049	0.053	0.355
Lymphoma	5	0.047	236.2	0.325	0.288	0.341	−0.013	0.068	0.845
Melanoma	45	0.026	244.37	−0.035	0.053	0.507	−0.002	0.032	0.950
Non-glioblastoma glioma/Glioma	19	0.052	88.03	0.102	0.049	0.052	−0.021	0.023	0.356
Oral cavity and pharyngeal cancer	4	0.059	198.2	0.008	0.376	0.986	0.094	0.064	0.144
Pancreatic cancer	16	0.037	68.9	−0.221	0.152	0.168	0.003	0.041	0.934
Prostate cancer	74	0.02	38.9	−0.091	0.060	0.130	−0.022	0.028	0.443
Renal cell carcinoma	8	0.028	148.02	−0.145	0.241	0.569	−0.031	0.084	0.707
Uterine fibroids	18	0.024	732.5	0.164	0.185	0.388	−0.014	0.073	0.854

PD, Parkinson’s disease; N, number; r², proportion of variance in exposure variable explained by SNPs; F, statistics ‘strength’ of the genetic instrumental variable; b, beta; se, standard error, p, p-value.

No causal effect of any cancer on PD was observed applying various MR methods (Table 1, Supplementary Table 1, Supplementary Figures 1–2).

To test for potential violations of MR assumptions, we performed sensitivity analyses. Significant heterogeneity was apparent for cutaneous squamous cell carcinoma (IVW, Q p-value = 0.02) and combined analysis of keratinocyte cancers (MR Egger, Q p-value = 0.012; IVW, Q p-value = 0.012, Supplementary Table 2, Supplementary Figure 3).

Tests for pleiotropy were performed to detect SNPs affecting the outcome through alternative pathways. There was some evidence for net horizontal pleiotropy for brain tumors (p = 0.011) and cutaneous squamous cell carcinoma (p = 0.029, Supplementary Table 2) which may have resulted in bias to IVW estimates, but the slopes from Egger regression were imprecisely estimated. Using MR-PRESSO, we detected an outlier SNP for cutaneous squamous cell carcinoma (rs4710154). The distortion test did not suggest significant changes in the effect estimates after this outlier was removed (Supplementary Table 2). The sensitivity analyses revealed no clear evidence for bias in the IVW estimate due to invalid instruments with other cancers.

Additionally, we performed reverse MR with melanoma, keratinocyte, prostate, endometrial and breast cancers for which we had full summary statistics using PD-associated SNPs as exposure and cancer summary statistics as outcome and did not find any evidence for causal relationships (Supplementary Table 3, Supplementary Figures 4–6). We found evidence for directional pleiotropy between PD and breast cancer and keratinocyte cancers, and a borderline distortion test with MR-PRESSO for breast cancer (Supplementary Table 3). MR-PRESSO identified an outlier SNP for both PD and breast and prostate cancer (rs4630591). Additionally, the rs510306 SNP was found to be an outlier for prostate cancer. For keratinocyte cancers, three outlier SNPs were detected (rs4630591, rs6599388 and rs4889603).

DISCUSSION

In the current study, we performed a comprehensive analysis to examine whether the reported associations between different cancers (Table 1) and PD may be causal. Our results provide no evidence to support causal effects and indicate that the observed associations may be due to other reasons including shared biology, confounders or biases. MR methods have limited availability and statistical power to differentiate horizontal and vertical pleiotropy, but high power to detect pleiotropy itself. Although MR can help reduce confounding and the possibility of reverse causality, a recent study demonstrated that MR is not immune to survival bias [31]. PD is an age-related disease and inverse observational study associations may occur spuriously if the exposure of interest (here cancer) causes premature mortality. This situation is known as ‘survivor bias’ and can occur in case-control settings, including in MR studies. On the other hand, early mortality from cancer could reduce cancer prevalence in PD [8]. The higher occurrence of brain cancers in PD might be related to closer medical attention (i.e., more frequent MRI in PD patients compared to the general population).

The most thoroughly studied genetic relationship between cancer and PD is for melanoma [32]. Previous MR studies did not demonstrate evidence of a causal relationship between PD and melanoma [22], However, a recent, comprehensive analysis suggested a significant genetic correlation between melanoma and PD, with gene expression overlap [10], that could probably explain the increased frequency of melanoma in PD. One of the possible explanations for the link between cancers and PD is pleiotropy. In our study, we only examined causality using MR and did not estimate possible shared biology. To study possible shared biology, methods such as linkage disequilibrium score regression and transcriptome wide association study can be used to examine correlations between two traits occurring through shared genetic architecture. Unfortunately, we were only able to collect full summary statistics of mostly sex-specific cancers (prostate, breast, endometrial cancers), which cannot be used with the PD GWAS data since it is not sex-stratified. This approach should be used in future studies. We cannot rule out that pleiotropic effects within the IVs cancel out each other if they have effects in opposite direction. There are genes involved in pathogenesis of both PD and cancers. It was suggested that familial PD genes (PINK1, DJI, LRRK2, etc.) may play a role in cancers [33–35]. GPNMB variants were associated with PD [36] and overexpression of GPNMB have been demonstrated in PD as well as in various cancers including melanoma [37, 38],

In our analyses using MR-PRESSO, we identified a few outlier SNPs. For cutaneous squamous cell carcinoma and PD, the rs4710154 SNP, located near the FGFRIOP gene, was an outlier. This gene was previously implicated in skin cancer and in several inflammatory disorders including Crohn’s disease [39]. This SNP was not previously associated with PD. Another outlier SNP, rs4630591, near the KANSL1 gene (encoding for KAT8 Regulatory NSL Complex Subunit 1) was identified for PD and breast and prostate cancers. This gene has been previously reported as the first cancer predisposition fusion gene [40], and this SNP was associated with breast cancer in transcriptome wide association study [41]. The rs510306 SNP near the IGSF9B gene has not been previously implicated in prostate cancer. For PD and keratinocyte cancers, three outlier SNPs were detected (rs4630591, rs6599388 and rs4889603). The rs6599388 SNP is located in TMEM175 and rs4889603 is located in STX1B, both of which have not been previously associated with skin cancers.

Our study has several limitations. This is a European-based study, and these associations or lack thereof should be studied in other populations. We excluded UKB data to decrease the chance of overlapping samples between studies, which can result in bias. As a result, some of our MR analyses might have not enough power to detect the causal effect. Lack of availability of sex-specific PD GWAS data is the another limitation, which would be important for studying the causal effect of sex-specific cancers, or with cancers that have meaningful sex differences [42]. We performed bi-directional MR with PD and cancers with available full summary statistics (melanoma, breast, prostate, endometrial and keratinocyte cancers) and did not find evidence of a causal relationships. One more limitation is that MR relies on the quality of the GWAS used for the MR, and thus, limited by the GWAS quality.

Additionally, we could not consider in the current analysis important environmental exposures that would be of interest for stratified analyses (e.g., smoking in lung cancer; hormone levels in sex-driven cancers). Thus, it is possible that we missed some causal effects due to gene-environment interaction or imperfect phenotype consideration.

To conclude, our results do not support a causal relationship between the tested cancers and PD and suggest that the observed associations could be a result of genetic pleiotropy, shared biology or biases. Once larger datasets become available, as well as sex-specific PD datasets, additional MR studies should be performed on cancers and PD.