The Effect of Circulating Zinc, Selenium, Copper and Vitamin K1 on COVID-19 Outcomes: A Mendelian Randomization Study

Abstract Background & Aims: Previous results from observational, interventional studies and in vitro experiments suggest that certain micronutrients possess anti-viral and immunomodulatory activities. In particular, it has been hypothesized that zinc, selenium, copper and vitamin K1 have strong potential for prophylaxis and treatment of COVID-19. We aimed to test whether genetically predicted Zn, Se, Cu or vitamin K1 levels have a causal effect on COVID-19 related outcomes, including risk of infection, hospitalization and critical illness. Methods: We employed a two-sample Mendelian Randomization (MR) analysis. Our genetic variants derived from European-ancestry GWAS reflected circulating levels of Zn, Cu, Se in red blood cells as well as Se and vitamin K1 in serum/plasma. For the COVID-19 outcome GWAS, we used infection, hospitalization or critical illness. Our inverse-variance weighted (IVW) MR analysis was complemented by sensitivity analyses including a more liberal selection of variants at a genome-wide sub-significant threshold, MR-Egger and weighted median/mode tests. Results: Circulating micronutrient levels show limited evidence of association with COVID-19 infection, with the odds ratio [OR] ranging from 0.97 (95% CI: 0.87–1.08, p-value = 0.55) for zinc to 1.07 (95% CI: 1.00–1.14, p-value = 0.06)—i.e., no beneficial effect for copper was observed per 1 SD increase in exposure. Similarly minimal evidence was obtained for the hospitalization and critical illness outcomes with OR from 0.98 (95% CI: 0.87–1.09, p-value = 0.66) for vitamin K1 to 1.07 (95% CI: 0.88–1.29, p-value = 0.49) for copper, and from 0.93 (95% CI: 0.72–1.19, p-value = 0.55) for vitamin K1 to 1.21 (95% CI: 0.79–1.86, p-value = 0.39) for zinc, respectively. Conclusions: This study does not provide evidence that supplementation with zinc, selenium, copper or vitamin K1 can prevent SARS-CoV-2 infection, critical illness or hospitalization for COVID-19.


Introduction
Highly transmissible and virulent in at-risk groups, SARS-CoV-2, the causal agent for COVID-19, has been sweeping the globe since December 2019. Despite intensive research, there are few effective prophylactic and early stage therapeutic interventions for COVID-19, with the exception of vaccines [1,2]. However, worldwide vaccine distribution remains highly inequitable, with less than 4% of the African population vaccinated as of September 2021 [3].
Corollary to drug repurposing efforts, the potential importance of micronutrients and their supplementation in preventing and alleviating COVID-19 has been proposed [4,5]. Supplements have some unique advantages as they are inexpensive, widely available overthe-counter, easily distributed and stored, generally well tolerated and well characterised in terms of safety. However, unlike drugs, the quality assurance of dietary supplements is not well-regulated, which can lead to risks associated with contamination, adulteration, fraudulent ingredients [6] and an incorrect, potentially toxic dosage [7,8]. Amongst vitamins and minerals, good mechanistic reasons for more research exist for zinc, selenium, copper in MR [48]. This is especially important as any observational studies linking nutrients levels to COVID-19 outcomes are confounded by fact that COVID-19 at-risk groups (e.g., the elderly, high BMI individuals, diabetics [49]) have, on average, lower/suboptimal levels of many micronutrients [13,15,18,44,46,[50][51][52] and, at the same time, suffer from poorer COVID-19 outcomes. MR analyses can help to clarify the causal pathway in such cases. As such, MR has been successfully applied in nutritional epidemiology [53], including in studies using our micronutrients of interest as exposure [54][55][56][57][58][59][60]. Here, we apply the MR framework to test whether genetically predicted Zn, Se, Cu or vitamin K 1 levels are causally related to COVID-19 outcomes: risk of infection, hospitalization and critical illness.

GWAS Studies
We searched the literature, in the OpenGWAS [61] and GWAS Catalog [62], for genetic instruments associated with zinc, copper, selenium and vitamin K 1 levels in populations of European ancestry.
We evaluated genetic instruments from the published GWAS of zinc, copper and selenium content of erythrocytes in the Queensland Institute of Medical Research (QIMR) twin cohort and whole blood selenium in the Avon Longitudinal Study of Parents and Children (ALSPAC) cohort of pregnant women [63], measured using inductively coupled plasma mass spectrometry. The red blood concentrations of those trace elements generally represent the overall nutritional status well [64] and the total blood measurement is a standard biomarker [65]. GWAS was adjusted for the following covariates: the analysis batch, haemoglobin concentration and analytical QC data.
For selenium instruments, we used a published fixed-effects meta-analysis of toe-nail selenium concentration, measured using neutron activation analysis in four Europeanancestry US cohorts (Coronary Artery Risk Development in Young Adults, Johnston County Osteoarthritis Project, Nurses' Health Study, Health Professionals Follow-up Study) coanalysed with the QIMR & ALSPAC GWAS results [66]. Toe-nail Se GWAS was adjusted for the following covariates: age, smoking status, geography and top eigenvectors. Compared to the circulating selenium, toe-nail content reflects more long-term Se exposure.
Vitamin K instruments were derived from a GWAS of phylloquinone (vitamin K 1 ), the primary circulating form of vitamin K, for which the measurements were available in two European ancestry CHARGE cohorts [67], namely, the Framingham Offspring Consortium and the Health, Aging and Body Composition study. Phylloquinone measurements were obtained in plasma/serum using reverse-phase high-performance liquid chromatography followed by fluorometric detection. Vitamin K 1 concentration in plasma reflects recent intake [44]. The GWAS included the following covariates: age, sex and study-specific stratification, including population structure.
Further details regarding GWAS cohorts, sample collection, analysis and quality control are provided in the Supplementary Methods.
We do not have detailed information on supplementation in the exposure GWAS studies, other than for the toe-nail selenium analysis, and individuals with toenail Se concentrations > 2.0 µg/g were excluded, as this could "reflect exogenous contamination or considerable excess ingestion of Se supplements". Genetic variants used as instrumental variables can influence life-long absorption, distribution and excretion of environmentally available micronutrient concentrations, whether dietary or supplemental, but within a natural range. Any substantial supplementation among the subjects of exposure GWAS could potentially introduce a measurement error and thereby dilute the effect seen in MR analysis and is one limitation of our study design.
For each micronutrient phenotype, we clumped the instruments using PLINK ver. v1.90b4.1 [68] and 1000 Genomes European reference panel [69], at the threshold of r 2 < 0.05 and clumping distance of 10 Mbp. The final set of instruments was derived by combining one representative SNP from each clump selected based on lowest p-value and presence in the outcome dataset (Supplementary Figure S1).

Zinc Genetic Instruments
One of the two instruments for zinc was missing (rs1532423); however, we found a good proxy for the missing instrument using 1000 Genomes European population using the LDproxy web app [70]: rs2453868, situated 34,383 bp away with r 2 of 0.93 and D of 1 (correlated alleles: rs1532423 A = rs2453868 C , rs1532423 G = rs2453868 T ). Overall, rs2120019 and rs1532423 account for 4.6% of variance in red blood cell copper concentration (Supplementary Table S1).

Selenium Genetic Instruments
Out of the 12 genome-wide significant SNPs found at two loci in the meta-analysis, four of the SNPs survived LD-pruning. However, only two variants could be used as instruments (rs921943, rs6859667) as SNPs in the other two clumps (rs6586282, rs1789953, rs234709) or their proxies were not available in the COVID-19 GWAS. The two instruments in the meta-analysis accounted for only 2.25% of variance in the trait (Supplementary Table S1). The results of the meta-analysis were initially presented as Z-scores which we converted to betas using the formula in Taylor et al. (2016) [71].

Copper Genetic Instruments
Two genome-wide significant instruments were identified: rs1175550 and rs2769264; altogether they account for 4.6% of variance in red blood cell copper concentration (Supplementary Table S1).

Vitamin K 1 Genetic Instruments
Overall, 11 signals at 6 loci were detected in the GWAS discovery stage at p-value < 5 × 10 −5 , with none of the SNPs reaching genome-wide significance. We found four SNPs of concern which were removed from downstream processing (Supplementary Methods). The three variants retained: rs4645543, rs4852146, rs6862071 (Supplementary Table S1) jointly explained 3.06% of variance in circulating phylloquinone concentration.

Sensitivity Analyses Using Subsignificant Hits
In addition to the main analysis, where we included only genome-wide significant hits at p-value 5 × 10 −8 (with the exception of vitamin K 1 , for which no variants crossed the threshold), we ran sensitivity analyses including variants selected at a more liberal, sub-significant threshold of 5 × 10 −5 . This allowed us to enhance the statistical power of analysis at the increased risk of violation of core MR assumptions. For zinc (Supplementary  Table S2), we used 12 SNPs (total R 2 of 14%), while for selenium we only had access to results from individual cohorts: QIMR (15 SNPs, R 2 = 13.26%) and ALSPAC (12 SNPs, R 2 = 10.93%) and for copper we arrived at 7 SNPs (R 2 = 10.03%).

Independence MR Criterion
We calculated the variance in each exposure, explained by each set of instruments (R 2 ) and F-statistics using the formulas in Yarmolinsky et al. (2018) [57]. We did not observe any weak instrument bias, with the F-statistic ranging from 15 to 172 (Supplementary Tables S1 and S2).

Exclusion Restriction MR Criterion
Both PhenoScanner V2 [72] and OpenGWAS [61] were used to assess presence of horizontal pleiotropy among the candidate variants using default settings (Supplementary  Tables S3 and S4, Supplementary Methods).
In addition to this, we conducted a leave-one-out analysis in our sensitivity checks which should minimize any possible confounding introduced by individual SNPs asso-ciated with height and RBC traits, whenever possible. Next, we calculated Cochran's Q statistic and I 2 to look for signs of heterogeneity, also indicative of pleiotropy. Finally, MR-Egger is one of the MR methods which we employed and which can detect directional horizontal pleiotropy if the intercept significantly deviates from 0 [73].

Selection of Outcomes
The largest publicly available GWAS to date on COVID-19, provided in the COVID-19 Host Genetics Initiative release 5, was selected [74]. This fixed-effect meta-analysis contains up to 49,562 COVID-19 patients and 2 million controls from 46 studies across 19 countries, however, we used results obtained in 35 European-only cohorts. The outcomes available were very severe (critical) COVID-19 (vs. population), hospitalized (vs. SARS-CoV-2 infected but non-hospitalized with COVID-19 or vs. population) and SARS-CoV-2 infection (vs. population) (Supplementary Table S5). The covariates used in the GWAS analysis were age and sex.

Statistical Analysis
We used the online mRnd power calculator to explore the limits of our MR analysis [75]. All the MR analyses were performed using TwoSampleMR [76] and Mendelian-Randomization [77] R packages. Our main method was the inverse variance weighted (IVW) random-effects meta-analysis of causal effects of individual instruments, as it is the most efficient; however, the IVW is biased in cases of unbalanced pleiotropy [78]. We complemented IVW with analyses using other MR methods: MR-Egger [73], weighted median-based and mode-based estimator. MR-Egger relaxes the assumption of balanced horizontal pleiotropy at the cost of reduced power, whereas weighted median estimator is still valid if only min. Of the total variants, 50% meet the three main MR assumptions [48]; mode-based estimator is similarly robust to outliers while being more conservative.

Ethics Statement
This study used publicly available summary data and no original data collection was undertaken for this manuscript. Evidence of ethical approval for all of the included GWAS studies can be found in previous publications. Our investigation is in accordance with the ethical guidelines of the 1975 Declaration of Helsinki.

Power Analysis
In the main analysis, SARS-CoV-2 infection was the outcome with the greatest power due to the highest number of cases and controls in the outcome GWAS. The minimum detectable one-tailed odds ratio (OR) at an 80% power ranged from 0.91 for selenium to 0.93 for zinc and copper (Supplementary Table S6). Furthermore, a COVID-19 hospitalization analysis revealed lower power, ranging from OR of 0.82 for Se to 0.87 for Zn and Cu. Poor power was found for very severe COVID-19 (OR of 0.75 for Se to 0.81 for Zn and Cu). In order to increase the power in our sensitivity analyses, we used sets of genome-wide sub-significant (max p-value of 5 × 10 −5 ) variants for zinc, selenium and copper (Supplementary Table S7). In the sub-significant instrument analyses, minimum detectable one-tailed odds ratios at 80% power were 0.95-0.96 for SARS-CoV-2 infection and 0.87-0.89 for very severe (critical) COVID-19.

Vitamin K 1
Using a limited set of three vitamin K 1 genome-wide sub-significant (max p-value of 5 × 10 −5 ) instruments, we were not able to detect any strong effect of genetically predicted circulating vitamin K 1 on COVID-19 outcomes (Table 1, Figure 4). In the IVW analysis, vitamin K 1 increment (per natural log transformed nmol/L) associated with OR of 0.99 (95% CI: 0.93-1.05, p-value = 0.68) for SARS-CoV-2 infection. Next, both hospitalization (ver. population) and very severe COVID-19 showed a lower OR for increased vitamin K 1 exposure (OR = 0.98, 95% CI: 0.87-1.09, p-value = 0.66 and OR = 0.93, 95% CI: 0.72-1.19, p-value = 0.55, respectively) but the CI comfortably overlapped with OR on either side of 1. The confidence intervals for our sensitivity methods (MR-Egger, weighted median and mode) overlapped the IVW CI, although the OR point estimates sometimes differed in terms of direction of effect.

Pleiotropic Bias
We did not detected a significant heterogeneity of effects between our variants using Cochrane's Q and pleiotropy using MR-Egger intercept in the zinc, selenium and copper MR analysis (Table 1). However, for vitamin K 1 we did find directional pleiotropy using MR-Egger, with the intercept differing significantly from 0. Therefore, causal estimates using this set of instruments are biased and need to be interpreted cautiously.

Discussion
Our analyses attempted to elucidate the potential of zinc, selenium, copper and vitamin K 1 in prophylaxis and treatment of COVID-19 using MR techniques. However, we found little evidence of a causal association of genetically predicted micronutrient concentration on COVID-19 outcomes.
Zinc's many antiviral, immunomodulatory and anti-inflammatory functions have generated a lot of interest for its potential use in COVID-19 management [13,18,51,52,79]. Consequently, over 50 RCTs, including zinc, albeit typically used as adjuvant treatment or prophylactic, have commenced (ClinicalTrials.Gov, (accessed on 1 September 2021)). One of the first to publish results, The COVID A to Z trial, which tested the direct effect of high-dose zinc supplementation, found no reduction in symptom duration in outpatients and was terminated early on [80]. In terms of prophylaxis, no beneficial effect of zinc supplementation was reported among the 370,000 British users of the COVID-19 Symptom Study app [81]. However, small observational studies reported lower serum Zn levels as a predictor of illness severity [82][83][84][85]. As a caveat, all the real life data and MR analyses do not include the application of ionophores, which may be necessary for zinc's antiviral inhibition [19,20].
Moreover, a number of observational studies established a positive correlation between low serum selenium levels and COVID-19 severity and mortality [83,[86][87][88]]. An ecological study found a significant positive association between hair selenium concentration and COVID-19 recovery rate in different provinces of China [89] which was subsequently replicated using local soil selenium concentrations as the predictor variable [90].
In general, one method by which to partially reconcile our findings for selenium and zinc with general micronutrient deficit in hospitalized or severely ill COVID-19 patients found in small studies derives from previous observations in critically ill individuals. Therefore, the initial hypozincemia and hyposelenemia is thought to stem from diseasedriven inflammatory process (acute phase response), is found chiefly in plasma (but not, e.g., erythrocytes) [18,91,92] and recovers over time in survivors. However, this does not mean that pre-existing deficiencies will not have a compounded negative impact at this stage [52].
A recently published study suggested that Cu status is correlated with the survival status of COVID-19 patients [36]. In contrast, another study from Skalny et al. [88] reported that plasma copper levels and the Cu/Zn ratio increased in more severe disease. This could reflect the fact that copper and zinc are antagonistically absorbed [93] and high serum Cu/Zn ratio is a marker for infection, as zinc gets redistributed to liver in the acute phase of the infection [94].
Despite evidence supporting the role of coagulation modulation by vitamin K in COVID-19 severity and poorer outcomes among hospitalized COVID-19 patients with lower vitamin K status [44,45], we did not detect any effects of circulating phylloquinone on very severe COVID-19 or other outcomes.
We conducted multiple sensitivity analyses, involving different methods (e.g., MR-Egger) and instrument selection, which revealed consistent results. In general, the inclusion of pleiotropic variants in MR is likely to skew the results away from the null, so it is reassuring that we find no effects also in the analyses involving variants selected at a more liberal sub-significant p-value threshold [78] and in vitamin K 1 analyses showing directional pleiotropic bias.
MR studies have confirmed RCT findings for many known risk factors, such as blood pressure and low-density lipoproteins [95]. MR has also indicated no causal effect of high-density lipoproteins and C-reactive protein [96,97] on cardiovascular disease, which, had it been reported earlier, could have saved a lot of effort and the cost of developing failed drugs. MR evidence can therefore deliver considerable insight about the prospect of a therapeutic. Here, current MR analyses do not support causal pathway between Zn, Se, Cu, vitamin K 1 blood levels and COVID-19 outcomes.
Thanks to the MR framework's use of genetic instrumental variables, the possibility of confounding and reverse causality was limited. Lately, some concern regarding the ubiquity of the collider bias in epidemiological investigations of COVID-19 was voiced [98]. However, the use of general-population control without a known COVID-19 infection in the outcome GWAS serves to decrease the likelihood of collider bias emerging, while not biasing effect size estimates in GWAS sensitivity analyses. We also included a hospitalized versus non-hospitalized GWAS outcome but even in these analyses at a higher risk of collider bias, the results broadly agreed with those from hospitalized versus population outcome. However, the ascertainment bias in the GWAS due to differential case reporting and varying SARS-CoV-2 exposure levels by, e.g., socio-economic status remains difficult to account for.
Another potential form of bias affecting the MR is population stratification. Since we included only European-ancestry samples in all our analyses, we limited the potential of this bias to skew the results. However, there are no strong biological reasons as to why our conclusions should not be generalizable to other ancestries.
The main limitation of our study is the relatively low power to detect modest effects of micronutrient levels on COVID-19 hospitalization and severity due to the few reliable genetic instruments available for micronutrients of interest and limited number of cases in the COVID-19 GWAS. This could be improved in the future as better-powered GWAS for both the exposures (particularly for vitamin K 1 , where no instruments reached genomewide significance) and outcome become available. Since we used the same GWAS for instrument discovery and effect estimate, our analysis is likely to suffer from the "winners' curse" and that it increased the weak instrument bias that would pull the results towards the null [78], which again could be rectified if new micronutrient GWASs are released.
Another limitation is that MR methods only used model linear effects within the normal range of micronutrient concentration, so any potential non-linear U/J-shaped, threshold effects will not be correctly estimated. Individual-level data for both exposure and outcome in the same population sample are required for such an analysis [78]. Additionally, MR cannot answer the question as to whether specific subgroups, such as micronutrient deficient individuals, can benefit from supplementation. Furthermore, the phenotypes used in an MR analysis typically correspond to lifelong exposure and small changes in micronutrient concentration, which does not exactly mirror intensive, high-dose clinical interventions.

Conclusions
In conclusion, we found little evidence of the effect of genetically predicted zinc, selenium, copper or vitamin K 1 levels towards preventing infection with SARS-CoV-2, and disease progression, including hospitalization or developing very severe COVID-19. Similar MR findings were obtained for two other promising micronutrients, namely, vitamin C [99] and vitamin D [100,101], suggesting that the utility of dietary supplementation for general population in the COVID-19 pandemic may be limited.  Data Availability Statement: All exposure data described in the article is provided in the supplementary tables. All outcome data is available from COVID-19 HGI website: https://www.COVID1 9hg.org/results/r5/ (accessed on 1 September 2021). Code for statistical analyses is available on: https://github.com/marynias/COVID19 (accessed on 5 January 2022).