Unbiased metabolome screen links serum urate to risk of Alzheimer's disease

Alzheimer's disease (AD) is a progressive and fatal neurodegenerative disease caused by a combination of genetic and environmental risk factors. The serum metabolome refers to a set of small-molecules which are an important determinant of cellular health. We obtained genome-wide association study (GWAS) summary statistics for serum concentrations of 376 metabolites which were population matched with 2 GWAS studies of AD. For each metabolite we performed 2-sample MR (2SMR) using an inverse variance weighted (IVW) estimate for significance testing. After Bonferroni multiple testing correction one metabolite was causally linked to AD in both GWAS: serum urate. This result was supported by robust 2SMR measures and sensitivity analyses. We applied 2SMR to test for a causal relationship between serum urate and other neurodegenerative diseases: Parkinson disease (PD) and Amyotrophic lateral sclerosis (ALS). In ALS but not PD we identified a nominally significant link between serum urate and disease-risk, although in this case increased serum urate was protective. We conclude that serum urate is a modulator of risk for neurodegeneration. Our work has implications for the design of preventative interventions.


Fig. 1. Unbiased screen for the identification of potential metabolites related to AD. (A)
Overview of our study plan. In order to identify metabolites causally linked to AD we performed an unbiased screen of 376 metabolites by 2SMR. After stringent Bonferroni multiple testing correction we identified a causal link between serum urate and AD. This result was further evaluated by robust 2SMR tests and sensitivity analyses. Finally we tested for a link between serum urate and other neurodegenerative diseases. (B-C) QQ-plots of IVW p -values demonstrate that six and two metabolites were significant after Bonferroni multiple testing correction (red line) in the larger ( Kunkle et al., 2019 ) (B) and smaller ( Lambert et al., 2013 ) (C) AD GWAS respectively. ( Hemani et al., 2018 ). This avoids much of the selection bias and reverse causation  which can plague casecontrol studies and even prospective cohort studies. Standard epidemiological studies are often underpowered due to the requirement for measurement of both the exposure and disease in the same cohort ( Julian et al., 2021 ), whereas 2SMR enables the use of separate cohorts to measure an exposure and outcome which makes larger well-powered studies more feasible.
We have developed methods for large-scale unbiased (by which we mean hypothesis-free or agnostic) application of 2SMR across large numbers of exposures including the entire metabolome ( Boddy et al., 2022 ). Here, application of 2SMR in this way identified a causal effect of serum urate on risk for AD using 2 different GWAS. We have followed this up with robust measures, sensitivity analyses, and tests for reverse causation. We also show that there may be an effect of urate on risk of amyotrophic lateral sclerosis (ALS), another neurodegenerative disease. We offer important new insights into an area of controversy. Our approach is summarized in Fig. 1 A.

Exposure GWAS
Genetic instruments used to measure lifetime serum metabolite levels were derived from publicly available GWAS ( Kettunen et al., 2016 ;Shin et al., 2014 ) for 376 metabolites. The 2 source studies differed in sample size and methodology: Kettunen et al. em-ployed NMR spectroscopy to quantify metabolites in serum samples from up to 24,925 individuals ( Kettunen et al., 2016 ); whereas Shin et al. employed liquid-phase and gas chromatography coupled with tandem mass spectroscopy in serum from 7824 individuals ( Shin et al., 2014 ). The sample population for both studies was almost exclusively European; both studies included ∼20 0 0 individuals from the KORA cohort (Cooperative Health Research in the Region of Augsburg). The different methodologies employed were complementary: NMR is more easily applied in a high-throughput manner but sensitivity is lower than mass spectroscopy based quantification ( Marshall and Powers, 2017 ). A total of 16 metabolites (4% of the total number) where measured in both source studies, not including serum urate. Downstream analysis was supported by a GWAS specific to serum urate including 110,347 individuals of whom > 90% were European ( Köttgen et al., 2013 ). This study was a meta-analysis of > 70 studies and included population controls together with a smaller number ( ∼3%) of individuals who suffered gout, which is causally linked to elevated serum urate. Urate was quantified as uric acid via the "uricase method" in the large majority of cohorts; this relies on a measurable change in UV absorbance which occurs upon addition of uricase to a sample, and is dependent on the concentration of uric acid.

Outcome GWAS
Unbiased screening utilized 2 AD GWAS composed of 63,926 individuals ( Kunkle et al., 2019 ) and 54,162 individuals ( Lambert et al., 2013 ). Individuals within these studies were overlapping although the sample size and the precise population mix varied between studies. Both GWAS compared patients with a clinical or pathological diagnosis of AD to population controls and focused on Europeans. Only 434 controls from the KORA cohort were used in both the outcome and the exposure GWAS. Overlap between exposure and outcome GWAS can lead to bias but given the small proportion of overlapping individuals ( < 1%), in this instance we consider the effect to be negligible .

Two-sample Mendelian Randomization (2SMR)
In our unbiased metabolome screen, we reported the multiplicative random effects inverse variance weighted (IVW) ( Burgess et al., 2019 ) estimate of causal inference for all 2SMR tests because this carries the most statistical power and is more robust to heterogeneity than a fixed effects IVW ( Julian et al., 2021 ). The unbiased screen for metabolites was performed as described previously ( Boddy et al., 2022 ), including a Bonferroni correction for multiple testing, but with some modifications. Genetic instruments were initially selected with a conservative p -value cut-off ( p < 5E-8). When the cut-off p -value is too stringent it can diminish power; however, if weak instruments are included this increases the risk of confounding pleiotropy or introduction on non-informative instruments which may bias toward the null ( Julian et al., 2021 ). To avoid both of these scenarios we sequentially trialled larger p -value cut-offs (multiplied by a factor of 10) to determine the lowest p -value cut-off for which 5-50 SNPs were associated with the exposure after clumping. This was performed in an automated manner. Identified SNPs within a 10kb window were clumped for independence using a stringent cut-off of R 2 ≤ 0.001 within a European reference panel; where SNPs were in linkage disequilibrium (LD) those with the lowest p -value were retained. Where an exposure SNP was unavailable in the outcome dataset, a proxy with a high degree of LD (R 2 ≥ 0.9) was identified within a European reference population ( Machiela and Chanock, 2015 ). The ef-fects of SNPs on outcomes and exposures were harmonized in order to ensure that the beta values were signed with respect to the same alleles. For palindromic alleles, those with minor allele frequency (MAF) > 0.42 were omitted from the analysis in order to reduce the risk of errors due to strand issues ( Hartwig et al., 2016 ). After initial instrument selection we applied radial MR to detect and exclude outlier instruments ( Bowden et al., 2018b ). As before ( Boddy et al., 2022 ), in the final step we excluded tests with too few or too many SNPs in order to minimize p -value inflation. Briefly: To ensure that we did not include false positive results in our unbiased screen we measured the inflation factor ( λ) which is the ratio of the observed median p -value to the expected median p -value. Under the assumption that the majority of statistical tests will be non-significant then λ should ∼1. It is expected that the majority of metabolites are unrelated to AD risk and therefore this is a reasonable assumption. The optimal number of instrumental SNPs was tuned to set λ as close as possible to 1. For both AD GWAS this resulted in a maximum number of instrumental SNPs of 15 per test. In the smaller AD GWAS ( Lambert et al., 2013 ), we excluded tests including < 7 instrumental SNPs and in the larger AD GWAS ( Kunkle et al., 2019 ) we excluded tests including < 5 instrumental SNPs. The final number of tests was used to determine the Bonferroni corrected p -value; in the smaller AD GWAS ( Lambert et al., 2013 ) the number of tests was 104 and therefore the threshold p -value was 0.05/104 = 4.8E-4, and in the larger AD GWAS ( Kunkle et al., 2019 ) the number of tests was 157 and therefore the threshold p -value was 0.05/157 = 3.2E-4.
Significance was calculated using a IVW estimate ( Burgess et al., 2013 ) which is well powered providing there is not an excess of invalid SNPs ( Bowden et al., 2016 ). To increase confidence in the IVW results from our unbiased screen, we performed a series of robust MR measures and sensitivity analyses. MR measures such as the weighted median ( Bowden et al., 2016 ), weighted mode ( Hartwig et al., 2017 ), and MR Egger ( Slob and Burgess, n.d. ) are relatively robust to the presence of invalid SNPs. The strength of instrumental variables associated with the exposure was evaluated using the F-statistic, except where we were unable to obtain effect allele frequencies (EAF) which prohibited this calculation. An F-statistic of > 10 per instrumental SNP was taken to signify a sufficiently strong correlation with the exposure (Burgess et al., 2011 ). Cochran's Q test served to identify heterogeneity between instrumental SNPs ( Bowden et al., 2018a ). Radial MR was applied to detect and exclude outlier instruments ( Bowden et al., 2018b ). Finally, we performed a leave-one-out (LOO) analysis to determine if any single SNP(s) were exerting a disproportionate effect ( Burgess et al., 2019 ). The MR-Egger intercept test determines whether there is directional horizontal pleiotropy. Finally, the presence of reverse causation, that is, an outcome which impacts an exposure, can violate assumptions underlying 2SMR and therefore we explicitly tested for this by reversing the direction of the 2SMR test ( Richmond et al., 2014 ).
In this work, the entire MR analyses were performed through TwoSampleMR (version 0.5.6) and Mendelian Randomization (version 0.6.0), Radial MR (version 1.0) and R (version 4.1.3) packages, respectively.

Unbiased 2SMR analysis highlights serum urate as a risk factor for AD
In our previous work, we have used an unbiased 2SMR approach to successfully identify and validate metabolites linked to risk for ALS ( Boddy et al., 2022 ). We applied a similar approach here to identify metabolites linked to risk of AD. To increase con-fidence in our findings we applied our approach simultaneously in 2 different AD GWAS ( Kunkle et al., 2019 ;Lambert et al., 2013 ) including a stringent Bonferroni multiple testing correction in both analyses ( Methods ). Only one metabolite was causally linked to AD in both analyses after multiple testing correction: serum urate (Supplementary Table 1, Supplementary Table 2, Fig. 1

Measurement of serum urate in a different GWAS confirms a causal link to AD
Our unbiased analysis has suggested a causal link between the serum urate and AD using 2 different AD GWAS. As a further confirmation we repeated the analysis using a different GWAS for serum urate to derive instrumental SNPs. This GWAS was larger (n = 110,347) but from a more heterogeneous population ( Köttgen et al., 2013 ); we selected 21-25 SNPs at a p -value threshold of < 5 × 10 −8 ( Methods ). We confirmed a causal association with AD in both the smaller AD GWAS ( Lambert et al., 2013 ) Table 1

Sensitivity analyses support the causative association of serum urate with AD
Our IVW 2SMR tests revealed the evidence of a causal effect of urate on AD which was consistent across multiple datasets. However, the IVW test is particularly vulnerable to pleiotropic instruments; therefore we performed a series of sensitivity analyses to check for confounding of the IVW test by SNP pleiotropy. The calculated F-statistic was > 10 for each instrument suggesting that we had achieved adequate instrument strength ( Methods ) ( Table 1 ), although it was not possible to calculate the F-statistic for instruments derived from the Köttgen et al. study ( Methods ). Cochran's Q test did not reveal the evidence of instrument heterogeneity ( Table 1 ). Egger intercept p -values were greater than 0.05, suggesting that there was no statistically significant directional horizontal pleiotropy ( Table 1 ). The LOO analysis revealed that the test of the larger urate GWAS ( Shin et al., 2014 ) against the smaller AD GWAS ( Lambert et al., 2013 ) was sensitive to removal of 4 different instrumental SNPs ( Table 1 ) however, none of the other analyses were abrogated by removal of any single instrumental SNP, suggesting that this observation may be related to an underpowered test rather than instrument heterogeneity.
A causal effect of an outcome on an exposure can violate key assumptions underlying 2SMR ( Julian et al., 2021 ). Using both AD and both serum urate GWAS we did not discover evidence that AD was causally linked to serum urate, taking AD as the exposure and serum urate as the outcome ( Table 2 , Fig. 3 A-D); this is often referred to as bidirectional MR. An exception was the IVW test ultising the smallest AD GWAS ( Lambert et al., 2013 ) and the smallest serum urate GWAS ( Shin et al., 2014 ) (IVW p = 2.12E-3) however, this result was not significant in any of the robust measures.

Robust 2SMR tests support the causative association of serum urate with AD
We have shown that the IVW test identifies a causal relationship between urate and AD and there is no evidence of confounding by reverse causation or instrument heterogeneity. To further increase confidence in our results we applied a series of robust MR measures which are less powerful than the IVW but less vulnerable to instrument heterogeneity.
Not all robust measures were found to be significant in MR robust tests. However, this is not unexpected given the relatively poor power of these measures ( Wang et al., 2020 ).

Evaluation of a causal effect of serum urate on risk of PD and ALS
After the validation of causal relation between urate and AD as supported by IVW and sensitivity/robust MR measures, we also tested whether serum urate was causally linked to other neurodegenerative diseases including PD or ALS by 2SMR ( Methods ). We observed that IVW results were not significant for the analysis of either serum urate GWAS ( Köttgen et al., 2013 ;Shin et al., 2014 ) against PD ( Nalls et al., 2019 ) (IVW p = 7.78E-1, beta = -0.016, se = 0.057, Fig. 4 C and IVW p = 8.00E-1, beta = -0.148, se = 0.586, Fig. 4 A; Table 3 ). For ALS GWAS ( Nicolas et al., 2018 ), there was some evidence of association between serum urate dataset measured using the smaller GWAS ( Shin et al., 2014 ) Table 4 , Fig. 4 B) and this test was also significant in the weighted median ( p = 0.021, beta = -1.007, se = 0.436, Table 4 ). However, sensitivity analyses revealed that this result was invalidated by removal of a single instrumental SNP ( Fig. 4 B, Table 4 ) suggesting that it may be a false positive. There was no evidence of association between the serum urate and risk of ALS using the larger urate GWAS ( Köttgen et al., 2013 ) ( Fig. 4 D,  Table 4 )

Discussion
In the present study, 2SMR was employed in an unbiased hypothesis-free fashion to explore the causal effect of serum metabolites on AD risk. Our results indicated that serum urate was causally associated with increased risk of AD. This result was supported by sensitivity analyses and robust MR measures.
This result was significant in 2 different AD GWAS after stringent multiple testing correction even when urate was considered alongside hundreds of other metabolites. Interestingly, we observed some evidence of a causal relationship between serum urate and risk of another neurodegenerative disease, ALS. However, in this case the direction of association was reversed and the sensitivity analysis suggested that this may be a false positive. If this result is confirmed in future works then this could indicate that serum urate is linked to neurodegeneration more broadly.
In our analyses we did not observe a simple relationship between the effect size and sample size as might be expected if adding more patients linearly increased statistical power to detect an effect. However, we note that larger sample size also increases population heterogeneity which can confound 2SMR because of a reliance on LD-structure for the selection of consistent instrumental SNPs ( Julian et al., 2021 ). Indeed we have previously discussed examples where the larger study does not necessarily provide the correct result ( Julian et al., 2021 ). It should also be noted that in the analysis of the largest urate GWAS against the largest AD GWAS the causal relationship between serum urate and risk of AD achieved significance in the IVW and both the weighted median and weighted mode robust 2SMR measures. Points indicate effect size ( β) and standard errors for each SNP-outcome relationship.  Urate is closely linked to uric acid (UA), a hydrophilic molecule which can scavenge free radicals including superoxide and singlet oxygen, that is essential for homeostasis within the CNS ( Bowman et al., 2010 ). Despite this, there is controversy regarding the association of urate and UA with neurodegenerative disease. Lower serum UA levels have been associated with AD ( Cankurtaran et al., 2013 ;Euser et al., 2009 ;Kim et al., 2006 ;Polidori and Mecocci, 2002 ) but another study found no significant differences in serum UA between aged healthy individuals and sufferers of AD or Parkinson's disease (PD) ( Ahlskog et al., 1995 ). A meta-analysis of multiple studies also concluded that serum UA levels in AD patients are not significantly different from healthy controls ( Chen et al., 2014 ). Higher serum UA has been proposed to reduce the rate of deterioration in cognitive function in individuals with mild cognitive impairment ( Irizarry et al., 2009 ;Rinaldi et al., 2003 ) but another study suggested that even a mild increase in UA level might increase the risk of cognitive decline in elderly adults ( Schretlen et al., 2007 ). Overall, controversy in the literature means that the role of urate and UA in AD suggests that there may be confounding, perhaps as a result of selection bias, which will not affect our 2SMR results ( Williams et al., 2019 ).
Other literature is consistent with our finding that urate is causally linked to AD and suggests a mechanism whereby UAinduced oxidative stress can exacerbate inflammation and formation of β-amyloid pathology. UA induces the generation of hydrogen peroxide ( Ko et al., 2019 ) and, causes the formation of superoxide anions when metabolized from xanthine in the UA biogenesis pathway ( Vannorsdall et al., 2014 ) suggesting that urate may contribute to oxidative stress in the cell ( Song and Zhao, 2018 ). This has been linked to core molecular defects underlying AD; a recent work concluded that long term exposure to serum urate might enhance the production of tumor necrosis factor-α (TNF-α) and βamyloid peptide resulting from oxidative stress in rats ( Tian et al., 2021 ).
A previous MR study of metabolites linked to risk of AD ( Lord et al., 2021 ) concluded that a number of lipid traits were causally associated with AD. Of note plasma urate was not considered in this hypothesis-led study. Similarly, our relatively conservative method in which we excluded traits with a large number of instrumental SNPs due to the risk of instrument pleiotropy, removed the large majority of lipid traits considered by Lord et al. (2021) . We propose that our findings could form the basis of therapeutic intervention, perhaps via dietary modification. A number of studies that have explored this, for example a longitudinal Taiwanese study concluded that risk of dementia in gout patients curated with urate-reducing agent was 30% less as compared to a control group or gout patients treated differently ( Hong et al., 2015 ). Similarly, dietary interventions thought to be associated with reduced serum urate, such as the Mediterraneantype diet, Dietary Approach to Stop Hypertension (DASH) and Mediterranean-DASH diet Intervention for Neurodegenerative Delay (MIND), have achieved reduction both in cognitive decline and AD incidence ( Solfrizzi et al., 2017 ). However, AD is associated with cardiovascular risk which may also have been modified by these dietary interventions.
We would like to draw attention to certain strengths and weaknesses of our work. Our analysis benefits from an unbiased datadriven approach to the serum metabolome, the use of multiple datasets, and our application of relatively stringent criteria (including robust tests and sensitivity analyses) for a positive result, all of which adds confidence to our findings. There is a degree of overlap in sample sets utilized by the 2 AD GWAS we employed ( Kunkle et al., 2019 ;Lambert et al., 2013 ) although, it is informative that despite this, serum urate was the only exposure to be causally associated in both datasets after Bonferroni multiple testing correction. A weakness of our work is the lack of biological interpretation because we have not performed any follow-up functional analyses. However, the starting point of this study was the metabolome and not specifically serum urate, and thus experimental evaluation of serum urate is beyond the scope of this work. It is notable that our 2SMR study was performed at population-scale which makes it difficult to make individualized predictions regarding the effect of serum urate on AD risk. Future work will need to identify which specific individuals or genetic backgrounds are at risk of urate-induced AD; these individuals may benefit from urate-lowering interventions.

Disclosure statement
All authors report no conflict of interest.

Supplementary materials
Supplementary material associated with this article can be found, in the online version, at doi: 10.1016/j.neurobiolaging.2022. 09.004 .