Increasing serum iron levels and their role in the risk of infectious diseases: a Mendelian randomization approach

Abstract

Objectives

Increased iron stores have been associated with elevated risks of different infectious diseases, suggesting that iron supplementation may increase the risk of infections. However, these associations may be biased by confounding or reverse causation. This is important, since up to 19% of the population takes iron supplementation. We used Mendelian randomization (MR) to bypass these biases and estimate the causal effect of iron on infections.

Methods

As instrumental variables, we used genetic variants associated with iron biomarkers in two genome-wide association studies (GWASs) of European ancestry participants. For outcomes, we used GWAS results from the UK Biobank, FinnGen, the COVID-19 Host Genetics Initiative or 23andMe, for seven infection phenotypes: ‘any infections’, combined, COVID-19 hospitalization, candidiasis, pneumonia, sepsis, skin and soft tissue infection (SSTI) and urinary tract infection (UTI).

Results

Most of our analyses showed increasing iron (measured by its biomarkers) was associated with only modest changes in the odds of infectious outcomes, with all 95% odds ratios confidence intervals within the 0.88 to 1.26 range. However, for the three predominantly bacterial infections (sepsis, SSTI, UTI), at least one analysis showed a nominally elevated risk with increased iron stores (P <0.05).

Conclusion

Using MR, we did not observe an increase in risk of most infectious diseases with increases in iron stores. However for bacterial infections, higher iron stores may increase odds of infections. Hence, using genetic variation in iron pathways as a proxy for iron supplementation, iron supplements are likely safe on a population level, but we should continue the current practice of conservative iron supplementation during bacterial infections or in those at high risk of developing them.

Key Messages.

• Iron levels have been associated with increased risk for some infections diseases.

• Iron supplement is a commonly prescribed medication in the population, and hence the effect of iron supplements on the risk of infections is unclear at a population level.

• Some randomized trials on iron supplements exist, but these do not translate well to the general population for which only limited evidence using potentially confounded observational studies exists.

• The use of genetic variants associated with higher iron reserves through Mendelian randomization can provide insight into the effect of iron supplementation on the risk of infectious diseases, while avoiding bias due to confounding and reverse causation.

• Increased iron reserves were not associated with a large difference in the risk of infectious diseases, though it was associated with a potentially worrisome increase in the risk of hospitalization from sepsis.

Introduction

Serum iron levels are modifiable through iron supplements and may strongly influence risk of infection.¹ Iron level and its related blood disorders are associated with multiple infectious diseases. For example, patients with hereditary haemochromatosis (a disorder of iron metabolism leading to a chronic iron overload state²) are at an increased risk of infections.³ Clinically we also frequently observe inflammatory anaemias in patients with infectious diseases, which is believed to be an adaptive response aimed at sequestrating iron stores away from infectious agents.⁴ Further, early studies on re-feeding individuals with nutritional deficiencies in remote communities affected by famine (e.g. nomadic people living near the Ethiopian-Somalian border in the 1980s⁵) supported an association between iron supplementation and risk of incident or reactivating infections, though these findings cannot easily be applied to other contexts. Hence, it is unclear if increasing iron intake through supplementation may put the general population at an elevated risk of infection. This has significant clinical implications, since up to 19% of the US population takes iron supplementation.⁶

Previously, traditional epidemiological studies on the role of iron in infection have involved subjects already at elevated risk of infections and who received iron supplementation or blood transfusions to treat the underlying diseases that already increased their risk of infections (e.g. patients receiving dialysis^7,8 or having a haematological malignancy⁹). Hence it is not clear how much iron levels contributed to their risk of infections and, unsurprisingly, these studies have reached varying conclusions¹⁰ Further, as ferritin (the standard test to measure a patient’s iron level status¹¹) is a known acute phase reactant that increases during infectious and inflammatory conditions⁴, monitoring iron reserves during supplementation in high infection risk individuals is likely to be biased due to inflammation due to infections (possibly indolent or subclinical). Hence, even if studying participants who are not otherwise at high risk of infectious diseases, reverse causation bias (where infections modify iron levels) could still lead to spurious causal associations. This could have important clinical significance, as a better understanding of the relationship between elevated serum iron levels and infection risks would lead to a reassessment of our iron supplementation treatment strategies. However, these biases make traditional epidemiological tools inadequate for studying this clinical question.

Mendelian randomization (MR) is a causal inference technique that in some ways is similar to a randomized trial, in that it uses genetic determinants of an exposure (here: iron) as statistical instruments to ‘randomize’ patients to different level of this exposure.¹² It reduces potential confounding between the exposure and the outcome (here: infectious disease) because of Mendel’s second law, which is the random assignment of alleles at conception, independently of nearly all variables that may be associated with infectious diseases. Further, it is not prone to reverse causality where the outcome can change the exposure, since genotypes are inherited at conception, always prior to disease onset. Its strengths have been illustrated on multiple occasions in a wide variety of settings, but most recently in the field of infectious diseases, MR studies have predicted randomized controlled trial results of interleukin-6 inhibition¹³, angiotensin-converting enzyme inhibition¹⁴ and vitamin D supplementation¹⁵ in COVID-19.

MR studies have three components:¹⁶ the outcome of interest (infectious disease), the exposure (iron status) that we hypothesize is causally related to the disease, and genetic markers for that exposure. It has three main assumptions: first, the genetic markers must be sufficiently associated with the exposure. Second, they should not be associated with any confounders of the exposure-outcome association. Third, their effect on the outcome is only mediated by their association with the exposure (also known as lack of horizontal pleiotropy). Specifically, this also assumes that the association between the exposure and the outcome is linear, though MR remains robust to non-linearity as it can still measure the population-averaged effect on the outcome when the entire distribution of the exposure is shifted in one direction.¹⁷ Nevertheless, the assumption of lack of horizontal pleiotropy remains the most problematic and requires careful adjustments and sensitivity analyses to ensure unbiased causal effect estimates.

Here we hypothesize that an individual’s iron status may affect their risk of developing infectious diseases. To test our hypothesis, we will combine genetic data from large biobank genome-wide association studies (GWASs) of iron and its biomarkers (ferritin, transferrin saturation and iron binding capacity) and multiple infectious diseases in an MR framework to obtain unbiased causal effect of iron on infections.

Methods

We used a two-sample MR design to estimate the causal effect of iron levels on the risk of infectious diseases. In a two-sample MR, the association between the genetic variants and the exposure and outcome are estimated in two independent cohorts.¹⁸ Assuming that these cohorts are epidemiologically and genetically similar, this allows the use of results from genome-wide significant associations in larger cohorts, thereby increasing sample size and statistical power. Compared with one-sample MR, it also has the benefit that any results we obtain will be biased towards the null hypothesis of no causal effect, rather than towards the estimate obtained from the epidemiological association (i.e. logistic regression of iron level on risk of infectious diseases) which suffer from confounding and reverse causation biases.¹⁸ A summary of our data source can be found in Table 1 below and in Supplementary Figure S1 (available as Supplementary data at IJE online). The study protocol was not pre-registered.

Table 1.

Data source

Exposure
Phenotype	Source	N		Number instruments with heterogeneity P-value <0.05a
Ferritin	Bell et al.20	246 293		6/48
	Canadian Longitudinal Study on Aging21	24 501
Iron	Bell et al.	163 511		–
TIBC	Bell et al.	135 430		–
TSAT	Bell et al.	131 471		–
Outcome
Phenotype	Source	N cases	N controls
Any infections	FinnGen30 January 2022 Release	70 078	190 327	5/58
	UK Biobank31	53 430	386 814
COVID-19 Hospitalization	COVID-19 Host Genetics Initiative Release 633,34	24 274	2 061 529	–
Candida infections	FinnGen January 2022 Release	2404	255 615	2/58
	UK Biobank	4882	435 362
Pneumonia	FinnGen January 2022 Release	33 723	226 682	5/58
	UK Biobank	21 518	418 726
	23andMe32	40 600	90 039
Sepsis	FinnGen January 2022 Release	7463	234 755	1/58
	UK Biobank	11 468	428 776
SSTI	FinnGen January 2022 Release	12 578	247 827	8/58
	UK Biobank	16 669	423 575
Urinary tract infections	UK Biobank	21 812	418 432	–

Summary of the sources of data used in our MR analyses.

TIBC, total iron binding capacity; TSAT, transferrin saturation; COVID-19 HGI, COVID-19 Host Genetics Initiative Hospitalization genomic-wide association study; SSTI, skin and soft tissue infection.

^*For meta-analysis of genome-wide association studies only.

Choice of iron genetic instruments

There are multiple ways to measure an individual’s iron reserve.¹⁹ These include measuring: (i) serum iron directly; (ii) ferritin, the protein that stores iron; (iii) total iron binding capacity (TIBC), a measurement of how much iron can be bound in blood; (iv) transferrin saturation (TSAT), a value obtained from dividing serum iron by TIBC and quantifying how much iron is being bound to transferrin, the main protein that binds iron in the circulation. Note that in usual physiological conditions, as iron stores change in the body, iron, ferritin and TSAT should be changing in the same direction, whereas TIBC would be going in the opposite direction. Again in usual physiological conditions, ferritin is considered the best measurement of overall iron reserves¹¹, though it does not represent the host’s ability to bind iron and ‘sequester’ it from infectious agents. In this study, we use independent genetic determinants of all four measurements obtained from the largest GWAS of iron biomarkers (to our knowledge) with measurements in up to 246 139 individuals of European genetic ancestry.²⁰ Full description of the distribution of ferritin, iron, TSAT and TIBC in this cohort are available in the original publication.²⁰ Note that only genetic instruments that were genome-wide significant for the given biomarker (P <5 x 10^-8) were used in analyses for that biomarker.

For ferritin specifically, we also performed a GWAS on ferritin in the Canadian Longitudinal Study on Aging²¹ (CLSA), an independent Canadian cohort for which ferritin values and whole-genome genotyping were available for 24 501 individuals of European genetic ancestry. The mean ferritin level in this cohort was 158, with a standard deviation of 142; full summary statistics on the distribution of ferritin levels in the CLSA are available in Supplementary Table S1 (available as Supplementary data at IJE online). Like Bell et al.²⁰, we first rank-based inverse normal transformed ferritin separately for each sex and we used the resulting values in our GWAS. The analysis was also adjusted for age, sex and the first 10 genetic principal components. This GWAS was done using the fastGWA²² function in the GCTA package (v1.93.2).

The resulting CLSA ferritin GWAS performed was then meta-analysed with an inverse-variance weighted fixed effect model to obtain the final ferritin GWAS which we extracted (with the metal package²³). Summary statistics harmonization prior to meta-analysis was done using EasyQC²⁴ (v23.8). After correcting for genomic inflation, we then used the linkage disequilibrium (LD) clumping algorithm to find lead independent genetic variants associated with ferritin, and these used in our MR analyses as statistical instruments for ferritin (implemented in plink^25,26 v1.9 with r² of 0.001 and window of 10 000 kilobases).

Last, for all analyses, we excluded genetic instruments falling within the major histocompatibility complex region of chromosome 6 (+/-500 kilobases). This locus is known for its complex linkage disequilibrium architecture and its involvement in multiple physiological processes. Hence, it is considered at high risk of horizontal pleiotropy, breaking one of the assumptions of MR. We also excluded any variants with minor allele frequency (MAF) of less than 0.5%, as these are at high risk of inaccurate imputation, even with state of the art imputation panels.²⁷

Strength of statistical instruments

We used the following approximation²⁸ to estimate the percentage of variance explained for each instrument (R²): where β and s.e. are the allele effect size on the exposure and its standard error, respectively. We then used the following formula for the F statistic of our instruments:

$$\frac{2\mathrm{*}{\beta }^2\mathrm{*}\mathit{MAF}\mathrm{*}\left(1-\mathit{MAF}\right)}{2\mathrm{*}{\beta }^2\mathrm{*}\left(\mathit{MAF}\right)\mathrm{*}\left(1-\mathit{MAF}\right)+{s.e.}^2\mathrm{*}2\mathrm{*}n\mathrm{*}\mathit{MAF}\mathrm{*}(1-\mathit{MAF})}$$

$$\left(\frac{n-k-1}{k}\right)\mathrm{*}\left(\frac{R^2}{1-R^2}\right)$$

where n is the sample size, and k is the number of instruments. An F statistic of more than 10 is usually sufficient to prevent weak instrument bias, though an F statistic of more than 100 is generally recommended.²⁹

Infectious diseases outcomes

We chose seven infectious diseases outcomes for our MR analyses (see Table 1): candida infections (also known as candidiasis), COVID-19 hospitalization, all-type pneumonia, skin and soft tissue infections (SSTI), sepsis, urinary tract infections (UTI) and having had recorded infectious diseases in medical records. These were chosen to have sufficient sample sizes to perform a sufficiently powered GWAS. When available, results from GWAS for all these conditions except for COVID-19 hospitalizations were obtained from three separate sources: the FinnGen Consortium³⁰, the UK Biobank³¹ and 23andMe.³² For FinnGen, we used publicly available summary statistics from GWAS of candidiasis, pneumonia, sepsis, SSTI and any recorded infectious diseases. FinnGen uses ICD-10 coding obtained from digital health records to define cases and controls. For 23andMe, we used results from a GWAS of pneumonia where cases and controls were defined based on client questionnaire self-reports. For COVID-19 hospitalization, we used summary statistics from the release 6 of the COVID-19 Host Genetics Initiative (COVID-19 HGI) worldwide meta-analysis.^33,34

For the UK Biobank, we used ICD-10 coding from hospital inpatient health-related outcomes (data fields 41202 and 41204) for defined cases and controls. ICD-10 codes were chosen to align with FinnGen case control definitions and were further curated by a certified physician. These can be found in Supplementary Table S2 (available as Supplementary data at IJE online). We then performed GWAS on each infectious disease using the Regenie software³⁵ (v2.2.4) to account for genetic relatedness between participants and to adjust for unbalanced case-control analyses using Firth logistic regression. Firth regression is a penalized likelihood method that provides unbiased effect size and standard error estimates even in highly unbalanced case-control situations. To prevent bias from population stratification, we restricted all analyses to individuals of European White British genetic ancestry.³⁶ The analyses were also adjusted for age at enrolment in the UK Biobank, reported sex, genotyping array and the first 10 genetic principal components.

Last, we meta-analysed the available GWAS from FinnGen, the UK Biobank and 23andMe (with the metal package²³), and used the results as outcomes for our MR analyses.

MR analyses

The effect of iron biomarkers on infectious diseases was estimated from every genetic instrument using the Wald ratio method. If a genetic instrument was not available in the outcome GWAS, or if it was palindromic and had an allele frequency >42% (making the chromosome strand ambiguous), we looked for a proxy variant with R² >95% in the 1000 Genomes Project³⁷ European ancestral populations. This was done using the MRutils toolkit and LDlink.^38,39 These were then meta-analysed using the inverse-variance weighted (IVW) and the MR-Egger methods. Whereas the IVW method is the most statistically powerful analysis, MR-Egger is robust to directional pleiotropic effect and provides an estimate of the amount pleiotropy with the alpha term (the intercept). All analyses were performed using the TwoSampleMR package⁴⁰ (v0.5.6) on R⁴¹ (v4.1.0), including variant harmonization.

Pleiotropy and heterogeneity sensitivity analyses

Last, to ensure that our results were not biased by horizontal pleiotropy, we performed the same MR analyses as above but after removal of genetic instruments at high risk of affecting the risk of infectious diseases other than through their effect on iron biomarkers. To choose these, we compiled a list of genes that are known to be directly involved in any of 11 gene ontology pathways involved in iron homeostasis (see Supplementary Table S3, available as Supplementary data at IJE online). We retrieved these genes and their positions using the Ensembl biomart⁴² online tool (Ensembl Genes version 105, GRCh38.p13 reference). This resulted in 85 genes whose main role is to maintain iron homeostasis and are therefore less likely to break the lack of horizontal pleiotropy MR assumption (see list in Supplementary Table S3). We then discarded any genetic instruments that were not within 500 kilobases of one of these genes.

We also used Cochran’s Q test of heterogeneity in causal estimates⁴³ to check robustness of our results. Finally, we performed MR using six additional methods that have different degrees of robustness to either pleiotropy or genetic instrument heterogeneity^20,43,44: MR Egger and the bootstrapped MR Egger, weighted median and weighted mode MR, simple mode MR and penalized weighted median MR. These were done on the full list of genetic instruments (with the exclusion of any genetic variants in the major histocompatibility complex).

Results

Choice of iron biomarkers genetic instruments

For iron, TIBC and TSAT, we used the linkage disequilibrium independent (R² <0.1) lead GWAS loci variants from Bell et al.²⁰ These resulted in nine genetic instruments for iron, nine genetic instruments for TIBC and 10 genetic instruments for TIBC. These variants explained a total of 2.5%, 3.2% and 2.9% of the genetic variance in iron, TIBC and TSAT, respectively. F statistics were 464 for iron, 504 for TIBC and 397 for TSAT, indicating appropriately powered instruments.²⁹ All variants were also found in the infectious GWAS (see below), none were palindromic with intermediate allele frequency and we did not need to look proxies in these analyses.

For ferritin, the LD clumping algorithm performed on the Bell et al.²⁰ and CLSA meta-analyses yielded 48 independent variants. Of those, rs255157 was palindromic with MAF of 43.6%, and was therefore replaced by rs255163. These explained 1.5% of the genetic variance in ferritin, for an F statistic of 88.6. However for the COVID-19 hospitalization phenotype, an additional four variants were not available in the outcome GWAS and no non-palindromic proxies with R² >95% could be found (rs1894692, rs192331981, rs190974828, rs149271440). The removal of these variants decreased the percentage of variance explained to 1.4% in this analysis, and decreased the F statistic to 87. A full list of genetic instruments summary statistics used for all analyses can be found in Supplementary Table S4 (available as Supplementary data at IJE online).

Infectious diseases GWAS and meta-analyses results

Sample sizes for infectious diseases case-control GWASs varied considerably but were generally comparable to other well-powered GWASs in the non-infectious diseases literature (Table 1). For example, the phenotype with the largest number of cases was the any-infections phenotype with 123 508 cases and 577 141 controls (from a combination of the UK Biobank and FinnGen), and the phenotype with the smallest number of cases was the candidiasis phenotype with 7286 cases and 690 977 controls (also from a combination of the UK Biobank and FinnGen). The pneumonia phenotype had the largest overall sample size with 95 841 cases and 735 447 controls (from UK Biobank, FinnGen and 23andMe). Manhattan plots and QQ-plots from each of our GWAS and meta-analyses results can be found in Supplementary Figure S2 (available as Supplementary data at IJE online), except for the COVID-19 hospitalization phenotype which has been reported elsewhere by the COVID-19 HGI.^33,34

Primary MR analyses

Using IVW meta-analysis to combine effect estimates from each genetic instruments, the only biomarkers associated with an elevated risk of infections were Iron and TSAT, for which one standard deviation increase led to a 1.10-fold increase in odds of sepsis (95% CI: 1.01, 1.20, P = 0.023) and a 1.08-fold increase, also in sepsis (95% CI: 1.01, 1.15, P = 0.016), respectively (Figure 1 and Table 2). Whereas these associations may be due to chance given multiple testing, it is reassuring that the effect direction of ferritin (increased risk of sepsis with increased ferritin) and of TIBC (increased risk of sepsis with decreased TIBC) were also in the expected direction.

Figure 1.

Mendelian randomization forest plots. Odds ratios for each infection corresponding to a one standard deviation increase in each iron biomarker (on the inverse rank normalized scale) and their corresponding 95% confidence intervals (CI). SSTI, skin and soft tissue infection; TIBC, total iron binding capacity; TSAT, transferrin saturation; UTI, urinary tract infection

Table 2.

Mendelian randomization primary analyses results

Outcome	Number of SNPs	IVW OR (95% CI)	IVW P-value	IVW SNP heterogeneity P-value	Egger alpha (95% CI)	Alpha P-value
Ferritin
Pneumonia	47	0.977 (0.913, 1.046)	0.511	3.51e-5	−0.0029 (-0.0083, 0.0024)	0.291
SSTI	47	1.045 (0.952, 1.146)	0.357	0.0151	0.0026 (-0.0048, 0.01)	0.488
Candidiasis	47	1.036 (0.853, 1.259)	0.719	0.00119	−0.0073 (-0.0228, 0.0082)	0.36
COVID-19 hospitalization	45	0.968 (0.824, 1.137)	0.692	2.42e-8	0.0122 (-4e-04, 0.0247)	0.064
Sepsis	47	1.09 (0.973, 1.22)	0.136	0.0201	0.0019 (-0.0071, 0.011)	0.675
Any infections	47	0.999 (0.955, 1.045)	0.963	0.188	−9e-04 (-0.0045, 0.0027)	0.625
UTI	47	1.017 (0.924, 1.119)	0.731	0.104	−0.0011 (-0.0088, 0.0066)	0.785
Iron
Pneumonia	9	0.999 (0.954, 1.047)	0.981	0.21	3e-04 (-0.008, 0.0087)	0.938
SSTI	9	1.097 (0.996, 1.208)	0.06	0.0239	−0.0076 (-0.0241, 0.009)	0.402
Candidiasis	9	1.058 (0.922, 1.214)	0.425	0.293	−0.0045 (-0.0293, 0.0203)	0.734
COVID-19 hospitalization	9	1.017 (0.882, 1.171)	0.819	0.00353	0.0133 (-0.0096, 0.0361)	0.293
Sepsis	9	1.103 (1.013, 1.2)	0.023	0.344	−0.003 (-0.0182, 0.0121)	0.708
Any infections	9	1.014 (0.972, 1.057)	0.526	0.192	−9e-04 (-0.0085, 0.0067)	0.832
UTI	9	1.026 (0.941, 1.119)	0.557	0.158	8e-04 (-0.0151, 0.0166)	0.926
TIBC
Pneumonia	9	1.019 (0.993, 1.046)	0.156	0.587	0.0029 (-0.0033, 0.0091)	0.393
SSTI	9	0.977 (0.906, 1.052)	0.532	0.00135	0.0068 (-0.0121, 0.0257)	0.503
Candidiasis	9	1.026 (0.94, 1.119)	0.567	0.328	0.0033 (-0.0192, 0.0258)	0.782
COVID-19 hospitalization	9	1.002 (0.94, 1.067)	0.959	0.193	0.0105 (-0.0032, 0.0242)	0.177
Sepsis	9	0.975 (0.901, 1.056)	0.536	0.0147	0.0065 (-0.0137, 0.0267)	0.549
Any infections	9	0.997 (0.967, 1.028)	0.841	0.0765	0.0071 (0.001, 0.0132)	0.056
UTI	9	0.987 (0.915, 1.064)	0.728	0.00663	0.013 (-0.0039, 0.0298)	0.176
TSAT
Pneumonia	10	0.99 (0.957, 1.023)	0.541	0.314	−0.0024 (-0.0087, 0.0038)	0.466
SSTI	10	1.069 (0.994, 1.15)	0.074	0.0219	−0.0131 (-0.0242, -0.002)	0.049
Candidiasis	10	1.015 (0.915, 1.126)	0.778	0.324	−0.0125 (-0.0308, 0.0059)	0.22
COVID-19 hospitalization	10	1.014 (0.911, 1.129)	0.797	0.00465	0.0079 (-0.0114, 0.0273)	0.445
Sepsis	10	1.078 (1.014, 1.147)	0.016	0.439	−0.0063 (-0.0175, 0.005)	0.309
Any i	10	1.01 (0.976, 1.045)	0.567	0.133	−0.0047 (-0.0104, 0.001)	0.146
UTI	10	1.024 (0.96, 1.093)	0.464	0.162	−0.0056 (-0.018, 0.0067)	0.397

Results of primary analyses and MR-Egger intercept terms. Odds ratios (OR) are given for an increase in one standard deviation in the iron biomarker (on the inverse rank normalized scale).

IVW, inverse-variance weighting; SNP, single nucleotide polymorphism; SSTI, skin and soft tissue infection; TIBC, total iron binding capacity; TSAT, transferrin saturation; UTI, urinary tract infection.

The other analyses did not show evidence for an effect of iron biomarkers on risks of infections. Importantly, the effect of iron biomarkers on the risk of admission with any infectious diseases was always close to the null, with the largest confidence interval found when using ferritin as the exposure [odds ratio (OR): 0.999, 95% CI: 0.955, 1.045, P = 0.96], suggesting that on a population level, increasing iron levels do not play a large role in the risk of infectious diseases.

Horizontal pleiotropy sensitivity analyses

Using the MR-Egger alpha intercept term, we did not find strong evidence of directional horizontal pleiotropy in any of our analyses (Table 2), with the largest alpha term found in the TSAT-SSTI analysis (alpha = −0.01, 95% CI: −0.02, −0.002, P = 0.049). Nevertheless, we also performed a sensitivity analysis with instruments restricted to genes in iron homeostasis pathways (Figure 2 and Table 3). Reassuringly for our previous analyses, the decrease in variance explained was mild except for ferritin, with percentage of variance explained decreasing to 0.8% for ferritin (from 1.5%), 2.4% for iron (from 2.5%), 3.2% (from 3.2%) for TIBC and 2.8% for TSAT (from 2.9%). Reassuringly, this likely mostly removed the weakest instrumental variants, as the F statistic increased to 139 for ferritin (from 88.6), to 499 for iron (from 464), to 635 for TIBC (from 504) and to 421 for TIBC (from 397). Effect estimates for sepsis were similar to our previous analyses but with larger confidence intervals, as expected, with a one standard deviation increase in iron associated with a 1.10-fold increase in odds of sepsis (95% CI: 0.99, 1.22, P = 0.064) and a one standard deviation increase in TSAT associated with a 1.07-fold increase in odds of sepsis (95% CI: 0.99, 1.16, P = 0.078). Further, we found an association between iron biomarkers and SSTI, with a one standard deviation in iron associated with a 1.10-increase in odds of SSTI (95% CI: 1.01, 1.21, P = 0.034), and a one standard deviation increase in TSAT associated with a 1.08-fold increase in odds of SSTI (95% CI: 1.01, 1.152, P = 0.023).

Figure 2.

Mendelian randomization sensitivity analyses forest plots. Odds ratios for each infection corresponding to a one standard deviation increase in each iron biomarker (on the inverse rank normalized scale) and their corresponding 95% CI. SSTI, skin and soft tissue infection; TIBC, total iron binding capacity; TSAT, transferrin saturation; UTI, urinary tract infection

Table 3.

Mendelian randomization sensitivity analyses results

Outcome	Number of SNPs	IVW OR (95% CI)	IVW P-value	IVW SNP heterogeneity P-value
Ferritin
Any infections	15	1.01 (0.941, 1.084)	0.784	0.123
Candidiasis	15	1.145 (0.853, 1.536)	0.368	0.0152
COVID-19 hospitalization	13	0.885 (0.742, 1.056)	0.175	0.132
Pneumonia	15	1.009 (0.89, 1.144)	0.891	1.28e-5
Sepsis	15	1.118 (0.952, 1.313)	0.173	9.8e-4
SSTI	15	1.01 (0.816, 1.249)	0.927	0.00553
UTI	15	1.017 (0.853, 1.213)	0. 849	0.00625
Iron
Any infections	8	1.017 (0.975, 1.061)	0.435	0.211
Candidiasis	8	1.068 (0.927, 1.231)	0.361	0.278
COVID-19 hospitalization	8	1.017 (0.872, 1.186)	0.833	0.00179
Pneumonia	8	1.004 (0.96, 1.05)	0.859	0.274
Sepsis	8	1.101 (0.994, 1.219)	0.064	0.256
SSTI	8	1.104 (1.008, 1.209)	0.034	0.0174
UTI	8	1.024 (0.932, 1.124)	0.622	0.11
TIBC
Any infections	7	0.997 (0.964, 1.031)	0.841	0.0501
Candidiasis	7	1.028 (0.943, 1.12)	0.531	0.355
COVID-19 hospitalization	7	1.003 (0.934, 1.078)	0.928	0.103
Pneumonia	7	1.017 (0.99, 1.043)	0.217	0.873
Sepsis	7	0.975 (0.895, 1.062)	0.563	0.0111
SSTI	7	0.973 (0.893, 1.06)	0.528	3.79e-4
UTI	7	0.988 (0.907, 1.077)	0.791	0.00204
TSAT
Any infections	9	1.012 (0.978, 1.048)	0.489	0.143
Candidiasis	9	1.021 (0.918, 1.136)	0.699	0.298
COVID-19 hospitalization	9	1.014 (0.904, 1.138)	0.811	0.00249
Pneumonia	9	0.993 (0.961, 1.025)	0.658	0.379
Sepsis	9	1.072 (0.992, 1.158)	0.078	0.345
SSTI	9	1.079 (1.011, 1.152)	0.023	0.0156
UTI	9	1.023 (0.954, 1.096)	0.522	0.116

Results of sensitivity analyses. ORs are given for an increase in one standard deviation in the iron biomarker (on the inverse rank normalized scale).

IVW, inverse-variance weighting; SSTI, skin and soft tissue infection; TIBC, total iron binding capacity; TSAT, transferrin saturation; UTI, urinary tract infection.

Genetic instrument heterogeneity

While there was no clear evidence in our sensitivity analysis, there was considerable heterogeneity in causal effect estimates in both the primary IVW MR analyses and the pleiotropy sensitivity analyses, especially for ferritin where the Cochran’s Q test for heterogeneity P-value was below 0.05 for most infectious diseases (Tables 2 and 3). We therefore used six additional MR analysis methods to check the robustness of our methods to both pleiotropy and heterogeneity. Whereas most of the analyses could not find a strong association, it is interesting to note that all analyses with P < 0.05 were found with TSAT as the exposure for either SSTI (largest OR: 1.15, 95% CI: 1.01, 1.26, P = 0.013), UTI (OR: 1.10, 95% CI: 1.02, 1.19, P = 0.019) or sepsis (OR: 1.12, 95% CI: 1.02, 1.22, P = 0.040) (Supplementary Figure S3 and Supplementary Table S5, available as Supplementary data at IJE online).

Discussion

In this MR study on iron levels and diverse infectious diseases phenotypes, we found a mixed picture of causal associations. Most importantly, we found that variations in iron levels are unlikely to be of sufficient magnitude to produce clinically relevant effects on infectious outcomes in the populations studied. However, elevated iron stores were likely associated with higher odds of infections in predominantly bacterial infections. Hence iron supplementation is likely safe in the large majority of patients, but decisions to supplement iron should still be made based on the patient’s individual infection risk profile.

First, we found that in most infectious phenotypes, if a causal association existed between increasing iron levels (as measure by an increase in ferritin, iron and TSAT or by a decrease in TIBC) and increased odds of infectious diseases, its overall population level effect would be mild to moderate. Indeed, in our primary IVW analysis (the most statistically powerful), all 95% confidence intervals for odds ratios were contained within the 0.882 to 1.259 range. Using genetic determinants of iron as proxies for iron supplementation, this suggests that in most individuals the risk of developing an infection due to iron supplements is likely to be balanced by its benefits, especially if they are not predisposed to infections due to other clinical factors (e.g. immunosuppression).

Second, although the overall picture drawn by our results is that increasing iron stores are unlikely to lead to a large increase in odds of infections, the three phenotypes for which at least one association was found were predominantly of bacterial aetiology (sepsis⁴⁵, SSTI⁴⁶ and UTI⁴⁷), whereas the other ones are either largely or definitely caused by viruses (COVID-19 hospitalization and pneumonia⁴⁸), caused by a fungus (candidiasis) or caused by any of the above families of infectious agents and parasites combined (‘any infections’ phenotype). These positive associations were also found mainly in TSAT, which along with ferritin is the key screening test for hereditary haemochromatosis⁴⁹, a Mendelian disease in which patients present with extreme levels of iron and which is associated with an elevated risk of certain bacterial infections.^50,51 However ferritin measurement is notoriously labile¹⁹, and our ferritin analysis may have suffered from weak instruments (see below), both biasing it toward the null hypothesis. Hence, our study supports current practice of conservative iron supplementation in patients with an active infection (especially bacterial) and may warrant that iron supplementation management be reviewed in patients with recurring infections or at high baseline risk of infections.

Third, it is interesting to note that whereas we found an association between iron levels and sepsis and UTI, the GWAS for these infections could only find one genome-wide significant variant (see Supplementary Figures S2.5 and S2.6, available as Supplementary data at IJE online). Given that their number of cases was much higher than for other phenotypes for which we found convincing GWAS associations, we suspect that this is due to their highly variable phenotype definitions (especially for sepsis). That is, both conditions can be caused by a wide range of infections which would probably deserve their own separate GWAS, and our GWAS were effectively testing for multiple diseases states at the same time. Hence, our results may suggest that for a subset of sepsis and UTI cases (e.g. only certain types of bacterial infections), higher iron level may be causing more infections. This shows the need for GWAS that are more specific to the different causal agents of sepsis and UTI, and also supports current clinical practice of conservative iron supplementation in patients with a high-risk infections or during active infections (at least until more research is done on the topic).

Our study has multiple limitations. First, our results cannot apply to individuals with extreme levels of iron storage. The cohort used in performing the GWAS for our MR analyses is meant to be representative of the general population (and may in fact be slightly healthier⁵²), and therefore our results should represent the overall population-level effect of genetically increased levels of iron storage on the risk of infectious diseases. This is even more true given the underlying assumption of linear exposure-outcome relationship.¹⁷ These results therefore do not represent the cases of individuals at the extreme ends of iron stores, i.e. patients with hereditary haemochromatosis with elevated iron levels or severely malnourished individuals with very low iron levels. Second, our study may suffer from weak instrument bias in which the genetic determinants of an exposure do not explain enough of its variance. In two-sample MR, this leads to a bias towards the null hypothesis of no causal effect.¹⁸ Although other studies have used MR to find causal effects with similar percentage of variance explained as ours and which were later corroborated by other lines of evidence⁵³, and F statistics used were all higher than 10, the ferritin analyses (with an F statistic of less than 100) may have been at higher risk of this bias. Third, iron supplementation could not be controlled for in most studies, which would also bias our results towards the null. Last, as we only studied the levels of iron biomarkers on individuals of European ancestry, it remains possible that iron may have different effects in other populations. However, to our knowledge there is no observational evidence to suggest that either iron deficiency or iron overload would affect risk of infectious diseases differently based on ancestry. This, once again, underlines the importance of developing large-scale genetic biobanks comprising people of non-European ancestry.

Conclusion

In conclusion, using MR, a method that accounts for the observational studies biases such as confounding and reverse causation, we find evidence that iron supplementation is unlikely to strongly increase the risk of infectious diseases in most of the population. However, in patients at high risk of bacterial infections, especially severe ones presenting with sepsis, iron supplementation strategy should likely remain conservative.

Ethics approval

All participants provided informed consent for analyses sharing of summary statistics and by their respective biobanks (see Table 1). All analyses and transfer of individual-patient level data was approved by the Jewish General Hospital Institutional Review Board.

Data availability

The UK Biobank, CLSA and 23andMe data are available through their respective data access process. Summary statistics from FinnGen are available freely on their website [https://www.finngen.fi/en]. Summary statistics from Bell et al.²⁰ are freely accessible through a link provided in their paper.

Supplementary data

Supplementary data are available at IJE online.

Author contributions

Concept and design: G.B.L., J.B.R. Data acquisition and standardization: G.B.L., Y.J., Y.C., T.N. Data analyses: G.B.L., Y.F. Interpretation: all authors. Computational resources and support: T.F., V.F., J.B.R. Writing original draft: G.B.L., J.B.R. All authors were involved in reviewing the manuscript and critically reviewed its content.

Funding

The Richards research group is supported by the Canadian Institutes of Health Research (CIHR: 365825; 409511), the Lady Davis Institute of the Jewish General Hospital, the Canadian Foundation for Innovation, the NIH Foundation, Cancer Research UK, Genome Québec, the Public Health Agency of Canada and the Fonds de Recherche Québec Santé (FRQS). T.N. is supported by Research Fellowships of the Japan Society for the Promotion of Science (JSPS) for Young Scientists and JSPS Overseas Challenge Program for Young Researchers. J.B.R. is supported by an FRQS Clinical Research Scholarship. Support from Calcul Québec and Compute Canada is acknowledged. TwinsUK is funded by the Wellcome Trust, Medical Research Council, European Union, National Institute for Health Research (NIHR)-funded BioResource, and Clinical Research Facility and Biomedical Research Centre based at Guy’s and St Thomas’ NHS Foundation Trust in partnership with King’s College London. These funding agencies had no role in the design, implementation or interpretation of this study.

Conflict of interest

J.B.R.’s institution has received investigator-initiated grant funding from Eli Lilly, GlaxoSmithKline and Biogen for projects unrelated to this research. He is the CEO of 5 Prime Sciences Inc. [www.5primesciences.com].