Circulating levels of micronutrients and risk of infections: a Mendelian randomization study

This study is reported according to the STROBE-MR (Additional file 1: Table S1) [16]. A schematic summary of the study design is given in Fig. 1. Briefly, we conducted a two-sample MR study using data from publicly available summary statistics from fourteen GWASs: eight for the exposures and six for the outcomes. Both exposure and outcome cohorts were restricted to subjects of European ancestry to reduce bias from population stratification [17]. All data used in this work are publicly available from studies with relevant participant consent and ethical approval, and ethical approval from an institutional review board was therefore not necessary for the present study.

We searched for published GWASs evaluating individuals of European ancestry on the GWAS Catalog and PubMed (the last search was performed in May 2022). We did not find any GWAS conducted for vitamins B1, B2, B3, B5, B7, sulfur, iodine, chloride, and fluoride. The GWASs conducted for vitamin K, potassium, sodium, cobalt, chromium, and molybdenum were excluded because of no significant genome-wide results [8, 18, 19]. In total, fourteen micronutrients of potential interest were identified: calcium [5], copper [6], iron [7], magnesium [8], selenium [6], zinc [6], beta-carotene [9], folate [10], vitamin A [11], vitamin B6 [12], vitamin B12 [10], vitamin C [13], vitamin D [14], and vitamin E [15] (Additional file 2: Additional Text) [5‐15, 20, 21]. For copper, we also identified a more recent and larger GWAS by Jäger et al. [20], but given that this study reported Z-scores and not beta-coefficients, we used the study by Evans et al. [6] in order to improve interpretability. However, the genetic instruments from the GWAS by Jäger et al. [20] were used in secondary analyses. Vitamin A and vitamin E were excluded because those GWASs were adjusted for body mass index (BMI) [22] which might introduce collider bias if the genetic instruments of the exposure of interest also have an effect on BMI [23].

For the main MR analysis, we included independent SNPs (r² < 0.001 within 10,000-kb windows), strongly associated (P ≤ 5E−08) with the blood level of each micronutrient.

Based on the disease incidence and availability of published summary statistics, we evaluated the risk of the following three infections: gastrointestinal infections, pneumonia, and urinary tract infections. We used publicly available summary statistics from two independent cohorts of European ancestry: UK Biobank (UKBiobank HRC-imputed) [24] and FinnGen Release 6 [25, 26] (Table 1). The GWAS conducted using the UKBiobank HRC-imputed data was performed using SAIGE (a generalized mixed model association test that uses the saddlepoint approximation to account for case-control imbalance), adjusted for genetic relatedness, sex, birth year, and the first four principal components [24]. The GWASs conducted on the FinnGen dataset were analyzed using SAIGE and were adjusted for sex, age, first ten principal components, and genotyping batch [25].

In both the UK Biobank and FinnGen, cases and controls were defined based on International Classification of Diseases codes (10th revision) from hospital records (Additional file 2: Tables S2-S4) [27‐32]. For each infectious disease, summary statistics were meta-analyzed using a random-effects model in METAL (version 2011-03-25) [33] (Additional file 3: Table S5). The meta-analyses included genomic control to account for residual population stratification [33]. Additionally, we conducted Cochran’s Q statistical test, included in METAL [33], to assess the heterogeneity between the two cohorts for the genetic instruments used for the outcomes.

The strength of each genetic instrument was estimated using the F statistic: F = R²(N − 2)/(1 − R²), where R² equals the proportion of variance explained by the genetic instrument and N is the effective sample size of the GWAS for the SNP-micronutrient association [34]. The R² value was calculated using the formula 2 × MAF(1 − MAF)beta², where beta represents the effect estimate of the genetic variant in the exposure, measured in standard deviation (SD) units, and MAF represents the minor allele frequency [35] (Table 2). The effect allele frequency was not available for the genetic instruments used for copper, selenium, and zinc, published by Evans et al. [6]. However, for the main analysis, none of the copper or selenium-associated SNPs were palindromic (A/T or G/C alleles), so it was clear which allele was the effect allele. For zinc, we removed one genetic instrument, rs10931753, due to being palindromic. The effect allele frequencies from the meta-analysis of FinnGen and UK Biobank were used to estimate the F statistics and R² for copper, selenium, and zinc. Although this may result in incorrect calculations of R² (and thus F statistic and statistical power), our results aligned well with the reported R² from Evans et al. [6], with an R² of 5% for copper. We reported a lower R² value for selenium (R² of 2.4%), but we only used 1 SNP, while Evans et al. [6] reported a total R² for 2 SNPs (R² of 4%) and a lower R² value for zinc (R² of 4.25%), where we used 2 SNPs, while Evans et al. [6] reported a total R² for 3 SNPs (R² of 8%). Additionally, for the secondary analysis of copper, we observed an unlikely R² which was explained by one extremely outlying SNP: rs12582659 (Additional file 2: Fig. S1). After removing this SNP, we found an R² of 7.10%.

Power calculations were done using http://​cnsgenomics.​com/​shiny/​mRnd/​ [36]. The statistical power was calculated to capture an odds ratio (OR) of 0.90 or 1.10 per SD change in the circulating micronutrient concentration, given the sample size used for the meta-analyses at a type 1 error of 5% (Additional file 2: Table S6). In addition to the main MR analyses, we performed secondary analyses using more liberal criteria for including genetic variants to enhance statistical power; r² < 0.01 and P ≤ 5E−06. For our study, we only considered micronutrients with an R² > 1% and/or statistical power > 50% for at least one of the infectious disease outcomes, thereby excluding calcium, magnesium, folate, and vitamin B6 (Table 2 and Additional file 2: Tables S6-S7). For the remaining micronutrients, we excluded SNPs with an F statistic < 10 [4] to reduce the risk of weak instrument bias [34]. None of the included genetic instruments was shared by any of the considered micronutrients (Additional file 2: Table S7).

We calculated the Wald ratio for each SNP, defined as the SNP-outcome association divided by the SNP-exposure association [37]. When multiple SNPs were available for a micronutrient, we summarized the effect calculated by the Wald ratio using an inverse-variance weighted (IVW) analysis [38]. All reported associations correspond to an OR for the outcome per SD increase in the genetically predicted circulating concentrations of the micronutrient. The MR analyses were performed separately for the outcome GWASs from UK Biobank, FinnGen, and meta-analysis of the two cohorts. If not otherwise specified, the meta-analysis was used as the outcome study. P < 0.05 was considered nominally significant, whereas the level for statistical significance corrected for multiple testing (8 exposures × 3 outcomes = 24 tests) was set at P = 0.05/24 = 2.08E−03.

For an instrumental variable to be valid, three key assumptions must be met: the instrument must be robustly associated with the exposure, it cannot affect a confounder of the exposure-outcome association, and it must only affect the outcome through the risk factor [4]. Horizontal pleiotropy—that the SNP has multiple effects—can violate those assumptions. MR-Egger, weighted median, simple mode, and weighted mode are some of the most common sensitivity analyses to account for horizontal pleiotropy [17]. These analyses were only conducted when the number of genetic instruments was ≥ 3. The MR-Egger method allows some SNPs to affect the outcome through a pathway other than through the exposure. If the intercept term differs from zero, this indicates that not all the included instruments are valid, and the standard estimates (i.e., IVW) may be biased [39]. The weighted median method provides a valid MR estimate when up to 50% of the included instruments are invalid. This method calculates the weighted median estimate by ordering the genetic variants according to the magnitude of their estimates [40]. The mode-based methods (simple mode and weighted mode) assume that the most common causal effect is consistent with the true causal effect, allowing some instruments to be invalid without biasing the estimated causal effect [41].

To evaluate if the differences in the individual effect sizes among the genetic instruments may be related to pleiotropic effects rather than chance, we conducted Cochran’s Q statistical test [42]. This test was only conducted when two or more variants were available, and a P < 0.05 was considered significant in the test for heterogeneity.

We further evaluated whether the genetic instruments were associated with other phenotypes using PhenoScanner V2 [43], available at http://​www.​phenoscanner.​medschl.​cam.​ac.​uk/​ (accessed 30 October 2022). Additionally, we performed leave-one-out analyses for micronutrients containing > 2 SNPs. This was performed to examine the robustness of the IVW estimates and if any specific SNP drove the association (which could be due to pleiotropy) [4].

Multivariable MR was used to evaluate whether any of the identified phenotypes on PhenoScanner had introduced bias due to pleiotropy [17]. The genetic variants for the potentially pleiotropic phenotypes were collected from IEU OpenGWAS [44]. Based on the identified potentially pleiotropic pathways, we conducted one multivariable MR analysis of copper on the risk of gastrointestinal infections where we included erythrocyte count (GWAS identifier: ukb-d-30250_irnt) and hemoglobin concentration (GWAS identifier: ebi-a-GCST004615) in the analysis.

In secondary analyses, we included variants at a more liberal threshold of r² < 0.01 and P ≤ 5E−06. While this could increase statistical power, it also may increase the risk of violating the MR assumptions and introduce weak instrument bias. Therefore, we included MR-RAPS, an MR method for correcting for bias introduced by weak instruments, using robust adjusted profile scores [45]. For copper, we carried out the secondary analyses both using instruments from Evans et al. [6] (as in the main analysis) and using instruments from Jäger et al. [20].

Finally, to validate the association between copper and gastrointestinal infections (see the “Results” section), we conducted the following two post hoc analyses: first, we retrieved the results from an additional European GWAS on gastrointestinal infections conducted by Nudel et al. [46] and conducted MR analyses on this independent cohort. Only one of the two copper SNPs, rs2769264, was available from this study, and no reliable proxy for the other copper SNP was available (defined by r² > 0.9; using European ancestry in the LDproxy tool from National Cancer Institute LDlink [47]).

Second, to assess the possibility of reverse causation, we conducted MR analyses of the association between the genetically predicted risk of gastrointestinal infection on the genetically predicted blood levels of copper. We used the meta-analysis results on gastrointestinal infection as exposure in this analysis, including SNPs suggestively associated with gastrointestinal infection (r² < 0.01 within 10,000-kb windows, P ≤ 5E−06). For the outcome, we retrieved the copper association summary-level statistics from Evans et al. [6].

All MR analyses were conducted using the TwoSampleMR package (version 0.5.6) [42] in R (version 4.0.3). METAL (version 2011-03-25) [33] was used to perform the meta-analyses of the outcomes.

After correction for multiple testing, the only statistically significant micronutrient-infection association was that of genetically predicted blood levels of copper and risk of gastrointestinal infections (Fig. 2). One SD increase in genetically predicted blood levels of copper was associated with an OR of 0.91 (95% confidence interval [CI] 0.87 to 0.97, P = 1.38E−03), 0.89 (95% CI 0.80 to 0.98, P = 1.67E−02), and 0.93 (95% CI 0.87 to 0.99, P = 1.98E−02), in the meta-analysis, UK Biobank, and FinnGen, respectively (Fig. 2 and Additional file 2: Table S8). A nominally significant association was observed for both selenium and vitamin D on the risk of gastrointestinal infections, with an OR of 0.92 (95% CI 0.85 to 0.99, P = 2.39E−02) and OR of 1.11 (95% CI 1.02 to 1.21, P = 2.00E−02), respectively (Fig. 2 and Additional file 2: Table S8).

We observed little evidence that the circulating concentrations of iron, zinc, beta-carotene, vitamin B12, and vitamin C were associated with the risk of any of the evaluated infections (Fig. 2 and Additional file 2: Tables S8-S10).

The MR-Egger, weighted median, and mode-based sensitivity analyses supported the findings from the IVW analyses. As only two instruments were used for copper, MR-Egger regression, weighted median, simple mode, and weighted mode were not carried out in the main analysis. No heterogeneity was observed in the main MR analyses for copper and vitamin D on the risk of gastrointestinal infections (Cochran’s Q test P = 5.99E−01 and Cochran’s Q test P = 5.56E−01, respectively; Additional file 2: Table S8). Heterogeneity was observed for vitamin B12 and gastrointestinal infections (Cochran’s Q test P = 2.19E−04): iron (Cochran’s Q test P = 1.04E−02) and vitamin C (Cochran’s Q test P = 3.70E−02) for pneumonia and vitamin D (Cochran’s Q test P = 2.27E−02) for urinary tract infection. For the other micronutrients, no heterogeneity was observed (Additional file 2: Tables S8-S10).

For the secondary analyses, we generally observed comparable effects as in the main analyses (Additional file 2: Table S11). However, for copper, using 6 SNPs from Evans et al. [6] yielded an IVW estimate of OR 1.01 (95% CI 0.96 to 1.06, P = 7.45E−01). There was considerable heterogeneity in this estimate (Cochran’s Q test P = 4.68E−03), which was explained by one extremely outlying SNP: rs12582659 (Additional file 2: Fig. S1). Excluding this SNP from the analysis yielded comparable results to the main analysis (OR 0.97, 95% CI 0.90 to 1.05, P = 4.35E−01; Additional file 2: Table S12). However, we still found evidence of heterogeneity (Cochran’s Q test P = 1.47E−02), and the results should be interpreted with caution. Using genetic instruments from Jäger et al. [20] supported the main analysis (OR 0.95, 95% CI 0.91 to 1.00, P = 3.07E−02). In both secondary analyses using data from Evans et al. [6] and Jäger et al. [20], we observed the presence of potential pleiotropy, e.g., MR-Egger OR 1.01 (95% CI 0.94 to 1.09, P = 7.18E−01) and OR 1.07 (95% CI 0.93 to 1.23, P = 3.99E−01); it is unclear whether this pleiotropy reflects the presence of pleiotropy in the main analysis.

Using PhenoScanner, we found that several of the genetic instruments used for the micronutrients have previously been reported to be associated with numerous traits and diseases (Additional file 3: Tables S13-S20). Of note, rs1175550 used for copper was strongly associated with reticulocyte count and hemoglobin concentration. We therefore performed a multivariable MR analysis between copper, reticulocyte count, and hemoglobin concentration and the risk of gastrointestinal infection. We observed a similar effect as in the main analysis with an OR of 0.96 (95% CI 0.93 to 0.99, P = 1.64E−02). The genetic instruments for vitamin D were associated with several traits related to smoking, BMI, and alcohol. In the leave-one-out analysis, the observed association of vitamin D with gastrointestinal infection did not change meaningfully, indicating that no specific SNP drove the result nor that the observed association was due to pleiotropy (Additional file 2: Table S21).

In the post hoc analysis of copper and gastrointestinal infections, we observed a similar effect as in the main analysis when using another GWAS on gastrointestinal infections [46]: OR 0.94 (95% CI 0.81 to 1.09, P = 3.82E−01). The CI was wide because only one of the two SNPs was available. Meta-analyzing the results from using Nudel et al. [46] with the main analysis yielded an OR of 0.92 (95% CI 0.87 to 0.97, P = 1.14E−03).

Finally, we conducted two-sample MR analyses using gastrointestinal infections as the exposure and blood levels of copper as the outcome. We found that gastrointestinal infections did not affect circulating copper levels (beta = − 0.35, 95% CI − 1.35 to 0.71, P = 5.40E−01) using two genetic instruments from Evans et al. [6], and no heterogeneity was observed (Cochran’s Q test P = 5.50E−01).