Dietary iron intake has long-term effects on the fecal metabolome and microbiome

Abstract

Iron is essential for life, but its imbalances can lead to severe health implications. Iron deficiency is the most common nutrient disorder worldwide, and iron dysregulation in early life has been found to cause long-lasting behavioral, cognitive, and neural effects. However, little is known about the effects of dietary iron on gut microbiome function and metabolism. In this study, we sought to investigate the impact of dietary iron on the fecal metabolome and microbiome by using mice fed with three diets with different iron content: an iron deficient, an iron sufficient (standard), and an iron overload diet for 7 weeks. Additionally, we sought to understand whether any observed changes would persist past the 7-week period of diet intervention. To assess this, all feeding groups were switched to a standard diet, and this feeding continued for an additional 7 weeks. Analysis of the fecal metabolome revealed that iron overload and deficiency significantly alter levels of peptides, nucleic acids, and lipids, including di- and tri-peptides containing branched-chain amino acids, inosine and guanosine, and several microbial conjugated bile acids. The observed changes in the fecal metabolome persist long after the switch back to a standard diet, with the cecal gut microbiota composition and function of each group distinct after the 7-week standard diet wash-out. Our results highlight the enduring metabolic consequences of nutritional imbalances, mediated by both the host and gut microbiome, which persist after returning to the original standard diets.

Introduction

Iron deficiency impacts more than two billion individuals worldwide, including ∼40% of the population in the developing world and ∼10% of the population in developed countries.^1,2 This makes it the most widespread nutritional deficiency across the globe. Deficiency of this essential metal can result in impaired immune response, stunted growth, and cognitive defects due to the essential role of iron in biological processes.^3–5 These include erythropoiesis, oxygen transport, mitochondrial respiration, electron transfer/mediation of oxidation-reduction reactions, hormone synthesis, DNA replication, and cell cycle control.⁶ Iron deficiency can occur during malnutrition but is also observed in overnutrition.^7–9 Furthermore, iron homeostasis is altered in metabolic diseases like inflammatory bowel disease and non-alcoholic fatty liver disease, in neurodegenerative diseases like Alzheimer's and Parkinson's diseases, and in cancers.^10–13 It is, therefore, essential to better understand how iron excess and deficiency can impact both host and microbial metabolism.

Nutrient availability profoundly shapes the composition and function of the microbial communities that colonize host organisms (humans, mice, etc.). This dynamic relationship is exceptionally relevant in the gut, where the levels of micronutrients, like polyphenols and flavonoids, or macronutrients, like carbohydrates and proteins, result in microbial communities with distinct compositions and functions.^14,15 Although underappreciated, metals, such as iron, copper, manganese, zinc, and cobalt, are also essential micronutrients required by all organisms and substantially influence the gut microbiota. Just as in higher organisms, iron is the most common redox-active metal found in enzymes and is a critical component of both iron-sulfur clusters and heme, co-factors required for ATP production, detoxification of toxins, and mediating oxidation-reduction reactions.¹⁶ Given these important roles for iron, host organisms use iron sequestration to control the growth of pathogens in a process called nutritional immunity, yet metal acquisition pathways remain understudied in the context of commensal human microorganisms.^17,18 For example, Bacteroides thetaiotaomicron, a human gut commensal bacterium, was recently found to use the siderophores enterobactin and salmochelin produced by other bacteria to survive in the inflamed gut.^19,20

Little is known about how the gut microbiota acquires iron and how dietary iron levels affect the function of these microorganisms, though studies have shown that dietary iron levels modulate microbial composition.^2,21 For example, a decades-old study found higher counts of Enterobacteriaceae in infants receiving iron-fortified cow milk than in those receiving unfortified breast milk.²² These results were reproduced in a study on the effects of dietary iron fortification in Kenyan infants, which also revealed increased levels of enterobacteria and decreased levels of lactobacilli.²³ These findings were dependent on diet, with a recent study showing that galacto-oligosaccharide/fructo-oligosaccharide prebiotics could modulate the gut microbiota toward a community rich in Bifidobacterium species.²⁴ Similar results have been observed in animal studies, with an increase in Enterobacteriaceae observed in both weaning pigs and rats fed iron-supplemented diets.^23,25 The study in rats investigated the effects of dietary iron deficiency and supplementation on short-chain fatty acid production but did not measure other metabolite levels.

While previous studies performed 16S rRNA sequencing to observe how changes in dietary iron levels affect gut microbiota composition, it remains unknown which enzymatic pathways of the gut microbial communities are modulated by these dietary changes.²⁵ Because of functional redundancy, where different microbial species encode for similar biological pathways, untargeted metabolomics and metagenomic sequencing are essential to understanding the functional response of the gut microbiome to changing dietary iron. Previous human and animal studies investigating the metabolome under altered iron intake have focused on plasma, blood, and liver samples, but effects on the fecal metabolome have been largely ignored.^26–30 Additionally, whether iron-mediated metabolic changes are reversible upon restoration of a standard diet remains an open question.

We performed a murine longitudinal study to address these knowledge gaps and assess whether dietary iron excess or deficiency can alter host and associated gut microbial metabolism in a way that persists past the dietary intervention period. Animals were assigned to one of three feeding groups: an iron-deficient, an iron-sufficient (standard), or an iron-overload diet. Ferric citrate, a commonly used oral iron supplement and an approved food ingredient by various regulatory boards, was added to supplement an iron-deficient diet to standard or excess iron levels.³¹ This iron source has been previously demonstrated to affect fecal iron levels and the abundance of bacterial taxa in mice.³² Moreover, it is well established that elevated levels of dietary iron promote hepcidin signaling in the liver, leading to reduced iron absorption and accumulation in the intestinal lumen. It is important to note that supplementation with ferric citrate may result in different outcomes than supplementation with ferrous sulfate due to the different oxidation states of iron and the inability of some bacteria to use ferric citrate as an iron source. Here, we investigate the impact of iron-deficient and overload diets on the murine fecal metabolome and microbiome. All three feeding groups were then switched back to an iron-sufficient diet after 7 weeks of dietary intervention and followed for an additional 7 weeks to assess whether observed metabolomic changes could be reversed by restoring standard iron levels.

Results

The fecal metabolome is shaped by dietary iron level and converges when standard dietary iron levels are restored

To investigate whether dietary iron influences the metabolism of the host and the gut microbiome during development, mice of mean age 32.6 days (± standard deviation of 2.7 days) were divided into three feeding groups: a Fe-deficient group (n = 5), a standard (Fe-sufficient) group (n = 5), and a Fe-overload group (n = 5). All 15 animals were fed a standard diet before the start of the study. This point was defined as study day 0 when the adolescent mice were divided into three groups and provided with the three different diets until study day 47. Fecal samples were collected at six time points (in addition to study day 0) before study day 47. After that, all three feeding groups were switched back to the same standard diet, and an additional six time points were collected (Fig. 1A). The fecal metabolome was characterized using untargeted liquid chromatography followed by tandem mass spectrometry (LC-MS/MS) on a Vanquish UHPLC coupled to a Q-Exactive orbitrap quadrupole mass spectrometer (Thermo Scientific) and feature detection in MZmine.⁷²

Fig. 1.

Dietary iron levels alter the fecal metabolome profile. (A) Design of the study. Mice (n = 15) were randomly assigned to three different diets, and they were housed separately in two cages per cohort. Days indicate fecal collection time points. The diet intervention started 30–36 days post-birth. (B) F statistic from PERMANOVA calculated at each timepoint during the study. (C) PCA score plots and PERMANOVA for study days 0, 47, 61, and 96. Group centroids are included as triangles and 95% confidence interval ellipses are included. Figure was created partially in Biorender.com.

The fecal metabolomes of the three diet groups started diverging at the first point of collection (study day 4) and continued to diverge for the remainder of the dietary intervention period (study day 47). Unsupervised principal component analysis (PCA) of the fecal metabolomes, along with permutational analysis of variance (PERMANOVA), was used to assess group separation (Fig. 1B). No separation of metabolomic profiles was observed among the three diet groups on the initial day of the study (day 0). Interestingly, divergence started on day 4 (F = 3.7, P = 0.001) and continued to increase until the end of the diet intervention at day 47 (F = 5.8, P = 0.002). While Fe-deficient and Fe-overload groups were switched back to standard, Fe-sufficient feed after day 47 fecal collection, the most significant difference between diet groups occurred at day 61 (F = 8.4, P = 0.001) (Figs. 1B and Fig. S1). At the final two time points (days 82 and 96), the three groups begin to converge, revealing that, eventually, the differences in the fecal metabolic profiles diminish with time after iron levels are restored. Finally, on day 96, the final time point for sample collection, the diet groups show obvious convergence of metabolic profiles. Four PCA score plots highlighting key time points in the study design are visualized in Fig. 1C. These time points were taken immediately before the start of the dietary intervention (day 0), the conclusion of the dietary intervention (day 47), the time point with maximal observed separation between groups (day 61), and the conclusion of the study (day 96). PCA was also used to investigate that no significant separation was driven by the individual animal (Fig. S2A) or cage (Fig. S2B). As expected, separation based on mouse age was observed (Fig. S2C). Taken together, unsupervised analysis suggests a strong relationship between dietary iron supplementation and the fecal metabolome of mice.

Differential dietary iron intake modulates lipid and peptide levels

To find metabolic features that differed between Fe-deficient, standard, and Fe-overload groups, pairwise partial least squares discriminant analysis (PLS-DA) models were built on the data collected on the final day of the dietary intervention (study day 47) (Fig. S3A, B). The top 30 features with the highest loadings and variable importance in projection (VIP) scores >1 from pairwise PLS-DA models for standard vs. Fe-deficient and standard vs. Fe-overload were extracted and plotted (Fig. S3C, D). In parallel, univariate statistical analysis, one-way ANOVA followed by Tukey's post-hoc test, was performed to corroborate the results of the supervised multivariate analyses. Each feature was either paired to a level 2 or level 3 annotation using the Global Natural Products Social molecular networking platform (GNPS) or an in silico annotation using SIRIUS and CANOPUS software.^33–37 This level system was proposed by the Metabolomics Standards Initiative (MSI) and has been further refined to communicate confidence in the structural assignment of MS features.^33,34 A level 2 annotation describes a probable structure, while a level 3 annotation describes a tentative candidate structure. While an exact structure can be proposed for a level 2 annotation, a level 3 annotation refers to the case where possible structures (i.e. positional isomers) can be proposed but one exact structure cannot be definitively assigned.

Twenty-five of the 30 features (83.3%) with the highest loading scores and VIP scores >1 from pairwise PLS-DA of Fe-deficient vs. standard diet were also significant in univariate analysis (Fig. 2A). The same was observed for the Fe-overload diet vs. the standard diet; 25 of the 30 features were significant in both multivariate and univariate analyses (Fig. 2B). The agreement between the supervised multivariate and univariate results led us to further investigate the molecular classes represented by the features driving the differences between the diet groups. Short asynchronous time-series analysis (SantaR) was used to investigate the fluctuation of the features with high loading scores across the duration of the study, and four of the 30 features are shown in Fig. 2C.³⁸ While some features were exclusively present in one diet group (oligopeptides in Fe-deficient diet; medium-chain fatty acids and purine nucleosides(t)ides in Fe-overload diet), other features were suppressed in only one diet, such as the methylpyridine family of metabolites in the case of Fe-overload. The temporal dynamics also differed based on feature; e.g. some features exhibited differences immediately on study day 4 (i.e. medium-chain fatty acid), while others showed longer delays in response (i.e. oligopeptide production). Using this analysis, we were able to prioritize molecular classes that were most impacted by dietary iron for further analysis.

Fig. 2.

Fecal oligopeptide, lipid, purine nucleoside, and methylpyridine molecular families are modulated by dietary iron levels. (A) Univariate analysis comparing standard and Fe-deficient diets shows consistency with the supervised analysis. Eleven out of 30 features extracted from the PLS-DA model are representative of oligopeptide molecular class. (B) Univariate analysis for standard vs. Fe-overload diets highlights presence of lipid molecules (mostly fatty-acids and derivatives; 7 out of 30) and purine nucleos(t)ide molecule in Fe-overload case. (C) Short asynchronous time-series analysis plots for selected significant features reveal changes in mean abundance over time. Short asynchronous time-series analysis plots were generated using feature abundances normalized to total ion count.

Fe-deficiency diet results in fecal di-, tri-, and oligopeptides increased abundance

Feature-based molecular networking (FBMN) was performed in GNPS on the full dataset across all collection days and visualized in Cytoscape 3.9.1 (Fig. S8).³⁹ In total, 6687 features were detected after blank subtraction, and among them, 374 were annotated in GNPS to yield a 6.5% annotation rate. Features were grouped by MS/MS spectral similarity, facilitating the organization of the large dataset obtained from the experiment. Visual assessment of the molecular network revealed that networks containing at least one annotated peptide member were predominantly present in the Fe-deficient diet samples. Moreover, given that numerous features with high predictive power for the Fe-deficient diet were annotated as peptides using in silico methods, we further investigated this molecular class.

Twenty features could be annotated as spectral matches to peptides in GNPS, with annotations concentrated as di- and tri-peptides.^33,34 Clear hierarchical clustering patterns emerge when heatmaps for annotated peptides were constructed (Fig. S10), with peptides exhibiting substantially higher abundance in Fe-deficient samples than in the other two dietary conditions. The increases in peptide abundance observed in the deficient vs. the normal diet begin around day 33, and notably continue to increase into the later days of the study (study days 68 and 75), as visualized in the short asynchronous time-series analysis (Fig. 3A). This pattern suggests either there is a lag in diet-associated changes or the switch back to the standard diet after the intervention period further modulates peptide metabolism.

Fig. 3.

Short asynchronous time-series analysis of molecular features modulated by dietary iron. Feature abundance normalized to total ion count was plotted as a time course analysis over all days of the diet study for the following molecular classes: (A) peptides, (B) fatty acids and derivatives, (C) purine nucleosides, and (D) conjugated bile acids. Each plot is constructed to compare two diet types (Fe-deficient vs. standard and Fe-overload vs. standard). Feature abundance for the omitted diet type matches the trend in the standard diet. Plots visualizing all three diet types at a time can be found in the supplementary information. The inset of each plot represents the molecular network to which the feature belongs and was retrieved from GNPS and visualized in Cytoscape.

Peptides consistently exhibited distinct profiles for the Fe-deficient diet compared to the other two diets. Peptide abundance increased at study day 33 and peaked around study day 68 before gradually returning to low levels that matched the standard diet (Figs. 3A and S9). This observation is intriguing for several reasons: first, the change in peptide profiling is reversible once iron is reintroduced into the deficient diet; and second, there appears to be a delay between the diet switch and the response in peptide abundance to this change. As with the peptide spectral matches, the same trend was also observed for the 743 features with in silico di- and tri-peptide annotations (Fig. S11). Heatmaps revealed a robust hierarchical clustering pattern, along with increased abundance in the Fe-deficient samples for most of these features during the later timepoints (Figs. S10 and S11).

Fe-overload diet results in increased fecal abundance of fatty acid derivatives and purine nucleosides

While di- and tripeptides were found to differentiate between standard and Fe-deficient diets, 13 of the 30 VIP features from PLS-DA of Fe-overload vs. standard diet were fatty acids and derivatives. Stearidonic acid and methionine-conjugated caproic acid level 3 annotations were among the features with increased abundance in Fe-overload (Fig. 3B). In both cases, the highest abundance was observed after the diet switch then decreased to levels present in the standard diet by study day 96. Several additional phospholipids, including specific lysophosphatidylcholines, exhibited distinct profiles in the Fe-overload diet. At the same time, certain phosphatidylethanolamine molecules were abundant in both standard and Fe-deficient diets but were below the detection limit in the Fe-overload diet. Distinct clustering of Fe-overload diet and Fe-deficient diet samples were also observed in heatmaps constructed for in silico annotated lipids (Fig. S13).

Along with fatty acids and their derivatives, purine nucleosides were also increased in abundance in the Fe-overload diet with respect to standard and Fe-deficient diets. Specifically, inosine, guanosine, and deoxyguanosine spectral matches were more abundant in Fe-overload (Fig. 3C). After the diet switched back to standard, all three molecules converged back to levels in the standard diet. This points to the reversibility of the iron-dependent increases in abundance. While the abundance of inosine in the Fe-overload diet began to increase on study day 33 of sample collection and peaked around study days 61–68, it eventually returned to levels matching those observed in the standard diet by the conclusion of the study (Fig. 3C). In contrast, increases in guanosine and deoxyguanosine abundance started earlier (day 4 for guanosine and day 19 for deoxyguanosine). Enrichment of guanosine and deoxyguanosine was consistently observed throughout most of the study period, even after the return to a standard diet, and only began to decrease on day 75. When the in silico annotated purine nucleosides and nucleotides (based on NPC class, ClassyFire most specific class, and ClassyFire class annotations) were clustered into a heatmap, Fe-overload samples exhibit a similar trend in the purine nucleoside abundance profiles (Fig. S15).^40,41

Iron intake levels influence microbial metabolite production

While peptides, fatty acid derivatives, and purine nucleosides can be produced by both the host and the gut microbiota, many amino acid-conjugated bile acids are produced exclusively by microbial communities.^42–44 This class of bile acid conjugates was also found to be influenced by dietary Fe levels. More broadly, bile acid homeostasis was altered in both Fe-overload and Fe-deficiency (Figs. 3D and S17). This finding was consistent with previous reports showing that Fe-overload in rats altered bile acid homeostasis through altered enzyme expression.⁴⁵ In addition to the host-produced bile acids reported previously, we found 12 of the 39 total spectral matches to conjugated bile acids were affected by differential iron supplementation (Fig. S16). Of these, nine trihydroxylated bile acids were found to be differentially abundant between the diets. These were primarily conjugated with aromatic and heterocyclic amino acids, including phenylalanine (Phe), tryptophan (Trp), tyrosine (Tyr), and histidine (His), as well as non-polar isoleucine and leucine (Ile/Leu), and polar threonine (Thr). The remaining three spectral matches were to recently discovered dihydroxylated bile acid conjugates, specifically conjugated to threonine (Thr), glutamate (Glu), and histidine (His).^42–44,46 While the distribution pattern of each conjugated bile acid varies, most detected bile acid conjugates exhibited reduced abundance in dietary Fe-overload compared to standard and Fe-deficient diets. An arginine-conjugated trihydroxylated bile acid was the only bile acid conjugate to exhibit a remarkably different trend, with a higher abundance in the Fe-deficient diet than in both standard and Fe-overload diets.⁴⁷ The abundance of this feature increased after the diet switch but reverted to standard diet levels by the end of the study.

Fecal metabolic profiles persist and eventually recover after the period of standard iron intake

To assess the persistence of observed metabolic changes for select molecular classes, we applied the following approach to the dataset from samples collected on study day 61 when the highest separation of metabolic profiles between three diets was observed (Fig. 1C). The top 30 features were extracted from each pairwise PLS-DA model comparing Fe-deficient vs. standard diets and Fe-overload vs. standard diets (Fig. S4), as performed previously for day 47. The significant extracted loadings for day 61 were compared to the significant features in univariate analysis. Notably, 26 out of 30 features (86.7%) exhibited significance in the univariate analysis for the Fe-deficient and standard diet comparison (Fig. S7A). 12 out of 30 features (40.0%) of differential loadings were significant for Fe-overload and standard diets (Fig. S7B). Inspecting spectral matches and in silico annotations of these features, consistent molecular class trends surfaced, resembling those observed during the dietary intervention period. Specifically, a heightened presence of oligopeptides was noted in the Fe-deficient diet compared to the standard group, while increased levels of fatty acids were evident in the Fe-overload group.

Most of the features significantly altered at day 47 from differential iron levels persisted into the dietary intervention period after day 47. Specifically, 26 of the 30 significant features for Fe-deficient vs. standard diet on day 47 were also significant (in the 100 features with the highest loading scores and VIP scores >1 from pairwise PLS-DA model in Fig. S4) on day 61, and 21 of the 30 significant features for Fe-overload vs. standard diet were also significant (in the 100 features with the highest loading scores and VIP scores >1 from pairwise PLS-DA model in Fig. S4) (Figs. S5, S6, Tables S2, S3). Twenty-eight of the 30 features for Fe-deficient vs. standard diet on day 47 have returned to levels found in the standard diet after the intervention period (Table S2), while all 30 features for Fe-overload vs. standard diet on day 47 have returned to levels found in the standard diet after the intervention period (Table S3), as assessed using Wilcoxon pairwise test with Benjamini-Hochberg false discovery rate (FDR) correction. This underscores the resilience and reversibility of the metabolic effects associated with different dietary iron loadings.

Iron intake influences cecal microbiome composition and function

The fecal metabolic profiles exhibited a lag between the conclusion of the dietary intervention (study day 47) and the day of maximal separation between groups through PCA (study day 61). Moreover, while the group separation decreases between study day 61 and study day 96, after the removal of the iron imbalance in their diet, there remains significant separation between the groups at study day 96 (PERMANOVA R² = 0.233, P = 0.001). We hypothesized that this persistence in the fecal metabolome difference was a result of changes in gut microbiota composition and functionality that took place alongside changes in host metabolism. To investigate this hypothesis, metagenomic sequencing was performed on cecal samples collected at the end of the study (study day 96). Unsupervised PCA of the robust center log ratio transformed count species and enzymes was consistent with this hypothesis, revealing significant separation between the three groups (PERMANOVA, R² = 0.24 and P = 0.014, Fig. 4A and C). Supervised PLS-DA confirmed unsupervised results [classification error rate (CER) <0.5] and differential abundance analysis using ALDEx2 was used to identify signature taxa and enzymes (Fig. 4B and D). The microbial taxa driving separations between the standard diet group and the Fe-deficient group were consistent with previously reported changes in the literature.^48–52 For example, species from both Lactobacillus and Limosilactobacillus genera were increased in abundance in the iron deficient group. This is consistent with previous reports that Limosilactobacillus and Lactobacillus both facilitate iron absorption during anemia.^49–51 On the other hand, species observed in higher abundance in the Fe-overload group, Anaerotruncus colihominis, Blautia coccoides, and Ruminococcus_B gnavus, have been previously associated with dysbiosis and high Fe-levels.^48,52

Fig. 4.

Dietary iron levels significantly alter cecal microbiome composition and functionality. (A) PCA reveals separation according to species between the three study groups at the end of the study (PERMANOVA, R² = 0.024, P = 0.014). (B) PLS-DA of cecal species profiles between the three study groups (CER = 0.35). (C) PCA reveals tendency in separation within encoded enzymes at the conclusion of the diet study (PERMANOVA, R² = 0.023, P = 0.076). (D) PLS-DA of enzymatic profiles between the three study groups (CER = 0.35). Group centroids are included as asterisks. ALDEx2 differential abundance analysis identified species signatures of (E) Fe-deficient vs. standard groups and (F) standard and Fe-overload groups and enzymatic signatures of (G) Fe-deficient and standard groups and (H) Fe-overload and standard groups.

In addition to the altered cecal microbial composition, differences in enzyme profiles between the diet groups were also observed. Differentially abundant enzymes with an effect size >|2| were further investigated, with the most differentially abundant enzymes shown in Fig. 4F–G. Interestingly, almost all enzymes were downregulated in cases of abnormal iron loading (both Fe-overload and Fe-deficiency) as compared to standard diet. Overall, the abundance of 39 enzymes was decreased in the Fe-deficient group compared to the standard one (Table S6), while the presence of 44 enzymes was decreased and 1 was increased in mice fed an Fe-overload as compared to a standard diet (Table S7).

The KEGG Mapper prediction tool was utilized to identify the pathways affected by altered dietary iron levels. Biosynthesis of secondary metabolites was the most affected pathway with 14 enzymes altered in this pathway.^53,54 Five enzymes in the purine metabolism pathway were also found to be affected in both Fe-deficient and Fe-overload conditions as compared to standard. Decreased presence of an additional enzyme (phosphoribosylglycinamide formyltransferase 1) was observed in Fe-deficiency as well as in Fe-overload (amidophosphoribosyltransferase). De novo purine biosynthesis was most affected by the modulation of dietary iron (Fig. S20) as the activity of key enzymes is iron-dependent. Five additional translocase enzymes linked to the hydrolysis of nucleoside triphosphates were also modulated by dietary iron. This observation is consistent with iron-dependent changes in purine nucleoside abundance observed in fecal metabolomics data (Figs. 3C and S14).

Branched-chain amino acid (BCAA) biosynthesis was altered in the Fe-deficient diet, and to a lesser degree, in Fe-overload diets (Fig. S21). Specifically, enzymes responsible for the first and third steps of BCAA biosynthesis, acetohydroxyacid synthase, and dihydroxy-acid dehydratase (DHAD), respectively, show high effect sizes in the Fe-deficient diet and less pronounced effects in the Fe-overload diet. DHAD contains two active-site Fe—S clusters that are likely modulated by iron availability.^55–58 In addition to BCAA biosynthesis, alanine (Ala), aspartate (Asp), and Glu biosynthesis, Phe, Tyr, and Trp biosynthesis along with Arg biosynthesis enzymes were present in lower abundance in both Fe-deficient and Fe-overload groups (Figs. S22–S24), while ornithine biosynthesis was specifically perturbed in the Fe-overload diet (Fig. S23). Fatty acid and polyunsaturated fatty acid biosynthesis were affected in the Fe-deficient diet vs. the standard diet. The gene encoding 3-oxoacyl-[acyl-carrier-protein] reductase was found at significantly lower levels in the Fe-deficient cohort compared to a standard diet. Altered lipid biosynthesis is consistent with metabolomics data, which revealed increased levels of fatty acids and PCs in Fe-overload diet compared to Fe-deficient and standard diets (Fig. S25).

Discussion

Differential dietary iron intake affected the fecal metabolome throughout the duration of the dietary study. Cecal microbiome composition and function were only assessed at the completion of the study, at which point composition and function still differed by dietary group. Observed changes in the metabolome and microbiome, which included altered levels of peptides, nucleic acids, and lipids (fatty acids and bile acids), persisted for 49 more days past the end of dietary intervention after Fe-deficient and Fe-overload feeding groups were switched back to a standard diet. These findings highlight the importance of dietary iron in both host and microbial metabolism and microbiome composition. The differential iron feeding altered levels of peptides, nucleic acids, and lipids (fatty acid and bile acid) metabolism.

These molecular classes were also consistent with Fe-dependent modulation of enzyme presence. For example, nearly half of the top 30 features extracted from the pairwise PLS-DA models between Fe-deficient and standard diets for the dietary intervention period were di- or tripeptides (Fig. 2A), and the abundance of enzymes involved in BCAA, Arg, Ala, Asp, and Glu metabolism was decreased in Fe-deficiency. These results indicate that encoding gene enzymes responsible for BCAA biosynthesis was reduced, possibly lowering de novo synthesis of amino acids. BCAAs can only be synthesized by bacteria and fungi but not animals, so reduced microbiome BCAA biosynthesis may be consistent with the prominent increases in di- and tri-peptides containing BCAA that were observed in the Fe-deficient group as compared to the standard diet (Fig. S10).⁵⁹ We speculate that protein catabolism by either the host or the gut microbiota may be responsible for the observed increase in di- and tripeptides as a compensatory mechanism for reduced biosynthesis.

Amongst the top features from the pairwise PLS-DA of Fe-overload and standard diets were mostly lipids (including both mono-, di-, and tri-carboxylic acids and bile acids) and nucleic acids. As in the Fe-deficient vs. standard pairwise comparison, the modulated metabolome features between Fe-overload vs. standard diet groups were consistent with altered enzyme presence at the study conclusion. For example, six enzymes in the purine metabolism pathway were more abundant in the Fe-overload group (Fig. S20), likely due to the iron dependence of key enzymes in this pathway. For example, the activity of ribonucleotide reductase, an enzyme essential for DNA synthesis and repair, depends on iron in its active site [4Fe-4S] cluster.^60–62 Enzymes participating in purine salvage, such as adenylosuccinate lyase (ADSL), also exhibit iron-dependent activity, with ADSL levels elevated in the liver of nursing piglets fed a high-iron diet.⁶³

Purine metabolites produced by both the host and gut microbiota serve as critical messengers beyond their canonical functions in DNA and RNA, and the gut microbial communities play a crucial role in modulating purine nucleoside levels.⁶⁴ Disruption of purinergic signaling has been linked to health conditions ranging from inflammation and neurodegeneration to cancer, as purines are crucial building blocks for nucleotides used by intestinal mucosa for immunity, energy production, and cell growth.^65,66 Moreover, increased purine salvage, triggered by the elevated degradation of purine nucleotides by both the host and the gut microbiota, has been described as contributing to inflammatory bowel disease, while gut bacterial purine catabolism may be a key mechanism driving atherosclerosis through the modulation of uric acid levels.^67,68 Consistent with this study, Fe-dependent modulation of purine metabolism was observed previously in liver and hippocampus metabolomes in a nursing piglet model; upon dietary Fe-overload treatment, guanosine, inosine, and other purine metabolites were detected to suggest a shift in flux from the salvage pathway toward degradation.⁶³ The present study provides the first evidence of the impact of dietary Fe on purine nucleosides in fecal samples, suggesting that dietary Fe affects a key aspect of gut microbiota-host signaling.

Bile acids have also recently emerged as important signaling molecules in the host that act through receptor binding.^43,69 Fe-overload in rats was previously demonstrated to alter bile acid homeostasis through reduced expression of Cyp7a1, the enzyme that performs the rate-limiting conversion of cholesterol into cholic acid, and decreased expression of Bsep, the transporter responsible for bile acid efflux.⁴⁵ This study was consistent with this previous report, as bile acid homeostasis was altered in both Fe-overload and Fe-deficient diet groups as compared to the standard diet group (Fig. S17). We were intrigued to find that amino acid conjugated bile acids were also modulated by dietary Fe levels. While recently discovered in biological samples, these molecules have reported bioactivities ranging from protection against the deleterious effects of high-fat diets to antagonizing the farnesoid X receptor.^42,43,70,71 Different conjugated bile acids exhibit specific patterns in Fe-deficient or Fe-supplemented diets as compared to standard diets. For example, Arg-conjugated trihydroxy bile acid increased during Fe-deficiency relative to the standard diet, while Glu-conjugated dihydroxy bile acid increased in the standard diet relative to Fe-overload (Fig. 3D). As unique microorganisms can produce each bile acid conjugate, we hypothesize that the different microbial taxa within each diet group result in a characteristic conjugated bile acid signature.^41,68

In conclusion, this is the first study demonstrating the role of dietary Fe in modulating both host and microbial metabolites in fecal samples. Importantly, this study also suggests that while many metabolites exhibit reversible abundance profiles when animals return to a standard diet, the gut microbiota composition and function remain divergent between the different diet groups even after a long wash-out period.

Limitations of the study

All MS/MS spectral matches are level 2 or 3 annotations, meaning that stereochemistry and regiochemistry may remain ambiguous. Additionally, a retention time drift of approximately 14 s was observed for the internal standard (Fig. S18), and although our feature finding in MZmine was set with a retention tolerance that takes this drift into account, there remain some split features. All spectral annotations with more than one match were manually inspected, and any split features (observed in MZmine) were summed. In these cases, this is explicitly stated. A limitation of the study design is that the distribution of male and female mice was not uniformly distributed; specifically, only female mice were fed the standard and high-iron diet, while both female and male mice were fed the iron-deficient diet.

Methods

Animals

Wild-type specific pathogen-free C57BL/6 mice were purchased from the Jackson Laboratory and used in our study, under protocols and guidelines approved by the Institutional Animal Care and Use Committee of the University of California, San Diego. Overall, each diet cohort contained five mice housed in two cages per cohort, with two to three mice per cage. In the standard and Fe-overload diet cohorts, there were five female mice housed in two cages per group. The Fe-deficient diet was fed to two male mice and three female mice housed in two separate cages.

Dietary intervention

The following diets were used in the study: (i) Iron deficiency: Teklad diet, TD.120514; (ii) Standard iron: in Teklad diet, TD.120515, 50 ppm ferrous sulfate was replaced with ferric citrate; (iii) Iron overload: Teklad diet, TD.120514, 6 g of ferric citrate iron/kg of diet (6000 ppm iron) were added.

Sample collection

Two hundred twenty-five total fecal pellets were collected. Fifteen fecal pellets were collected on the following days: days 0, 4, 8, 12, 19, 26, 33, 40, 47, 54, 61, 68, 75, 82, and 96. Mice were sacrificed on day 96 and organs were collected under sterile conditions.

Sample processing for MS

Fecal samples were thawed on ice for 30 mins before adding a stainless-steel bead to every sample. 80% MeOH solvent was added to each sample to maintain a mass-to-volume ratio of 1 mg per 10 µL. Samples were homogenized at 25 Hz for 5 min, centrifuged at max speed for 15 min, then supernatant was transferred and dried in vacuo overnight. Samples were stored at −80°C until analysis, at which point they were reconstituted in 80% MeOH + 1µM sulfadimethoxine to 1 mg/mL.

UHPLC-MS/MS

For LC-MS/MS analysis, 5 µL was injected into a Vanquish UHPLC system coupled to a Q-Exactive orbitrap mass spectrometer (Thermo Fisher Scientific, Bremen, Germany). For the chromatographic separation, a C₁₈ porous core column (Kinetex C₁₈, 50 × 1.0 mm, 1.8 µm particle size, 100 A pore size, Phenomenex, Torrance, USA) was used. For gradient elution, a high-pressure binary gradient system was used. The mobile phase consisted of solvent A H₂O + 0.1% formic acid (FA) and solvent B acetonitrile (ACN) + 0.1% FA. The flow rate was set to 0.15 mL/min. After injection, the samples were eluted with one of the following linear gradients: 0–1 min, 5% B, 1–7 min, 5–99% B, followed by a 2.5 min washout phase at 99% B and a 1.5 min re-equilibration phase at 5% B. Data dependent acquisition (DDA) of MS/MS spectra was performed in positive mode. Electrospray ionization parameters were set to 40 L/min sheath gas flow, 14 L/min auxiliary gas flow, 0 L/min sweep gas flow and 400°C auxiliary gas temperature; the spray voltage was set to 3.5 kV and the inlet capillary to 320°C and 50 V S-lens level was applied. The MS scan range was set to 150–1500 m/z with a resolution of m/z 200 (R_{m/z 200}) of 35 000 with one micro-scan. The maximum ion injection time was set to 100 ms with an automated gain control (AGC) target of 1.0E⁶. Up to five MS/MS spectra per MS1 survey scan were recorded in DDA mode with R_{m/z 200} of 17 500 with one micro-scan. The maximum ion injection time for MS/MS scans was set to 100 ms with an AGC target of 5E⁵ ions. The MS/MS precursor isolation window was set to m/z 1. Normalized collision energy was set to a stepwise increase from 20 to 30 to 40% with z = 1 as the default charge state. MS/MS scans were triggered at the apex of chromatographic peaks within 2–15 s from their first occurrence. Dynamic precursor exclusion was set to 5 s. Ions with unassigned charge states were excluded from MS/MS acquisition as well as isotope peaks.

Feature finding and molecular networking

Raw data conversion to mzML format and peak picking were performed using MSConvert and MZmine. Both .raw and .mzML files were uploaded to the MASSIVE database and made available for public access (ftp://massive.ucsd.edu/MSV000084783/). Data processing was carried out in MZmine 3.4⁷² using the parameters specified in the .xml batch file on the supplementary information. Subsequently, FBMN was performed in the GNPS platform using the .mgf file and feature table retrieved from MZmine 3.4, along with the metadata file. Bray-Curtis PCoA Distance Metric and row sum normalization were applied to the GNPS job. The GNPS job link can be found here: https://gnps.ucsd.edu/ProteoSAFe/status.jsp?task=d35b583c2cf3435289a0aca8de4df4dd.

Statistical analysis

The mass spectrometry data were further processed in the statistical software environment R (version 4.1.2). The MZmine feature table was blank subtracted with a cutoff value of 0.3 (i.e. features with a ratio of mean intensity in blanks vs. mean intensity of features less than 30% were considered background noise and removed). After blank subtraction, the minimum Limit of Detection imputation was applied. The imputed data table was then either normalized by the total ion count per sample or was subjected to centered log-ratio (CLR) transformation using the vegan 2.6.4 package. The imputed, normalized data were used in short asynchronous time-series analysis, while the CLR-scaled feature table was utilized in univariate and multivariate analyses.

The processed feature table was merged with the metadata table for the univariate analysis and checked for normality using the Shapiro-Wilk test. Due to the interest in the pairwise comparison of abnormal Fe loadings to standard dietary levels, we divided the data into different sets based on diet types and time: Standard & Fe-overload before the diet intervention, Standard & Fe-deficient before the diet intervention, Standard & Fe-overload after the wash-out, Standard & Fe-deficient after the wash-out. Data from each diet group was checked separately for normality before univariate analysis, and most features in every dataset were found to be normally distributed (Table S4). Therefore, one-way ANOVA was performed on each feature against the sample area followed by the Tukey HSD post-hoc test to evaluate the differences among different diets at a significance level of P < 0.05 after FDR correction.

Unsupervised PCA and supervised partial least squares-discriminant analysis (PLS-DA) were conducted to uncover significant variations between the different diet groups. PCA was performed using the factoextra 1.0.7 package on the data from separate days, as well as on the whole dataset. The results of the PERMANOVA analysis, performed with the adonis2 function from vegan package, were utilized to quantify the most significant separation for each independent time point (day of fecal sample collection). PLS-DA models were built using the mixOmics 6.22.0 package, where diet groups were assigned as response variables. The evaluation of PLS-DA model performance was carried out using leave-one-out cross validation. Discriminating metabolites were identified through VIP scores >1.

In silico chemical classification

From the initially identified 6687 features, only 606 compounds were successfully matched with public libraries for annotation. To enhance the identification of unknown compounds, annotation was additionally propagated by employing the in silico tools CANOPUS and SIRIUS.^36,37 The mgf result file from MZmine 3.4 as uploaded to the SIRIUS application, and specific parameters were selected, including the following: Database: none, Mass deviation of the fragment peaks in ppm: 5, Maximum number of candidates in the output: 10, Ion mode: positive, Analysis used: Orbitrap, Schema: Auto, Minimum number of MS/MS peaks: 1. As a result, an additional 5550 compounds were in silico annotated, propagating identification of previously unknown metabolites.

Data visualization

Molecular networking results obtained from GNPS were visualized in Cytoscape, with nodes represented as pie charts. The color distribution in each pie chart corresponded to the relative abundance of metabolites in each diet type.³⁹ To visualize the dose dependence of selected metabolites on dietary iron, the santaR 1.2.3 package was employed to plot feature abundance against time, and the gplots 3.1.3 package was used to generate heatmaps.

Whole genome sequencing

The UC San Diego Microbiome Core performed nucleic acid extractions utilizing previously published protocols.⁷³ Briefly, samples were purified using the MagMAX Microbiome Ultra Nucleic Acid Isolation Kit (Thermo Fisher Scientific, USA) and automated on KingFisher Flex robots (Thermo Fisher Scientific, USA). Blank controls and mock communities (Zymo Research Corporation, USA) were included and carried through all downstream processing steps. DNA was quantified using a PicoGreen fluorescence assay (Thermo Fisher Scientific, USA) and metagenomic libraries were prepared with KAPA HyperPlus kits (Roche Diagnostics, USA) following manufacturer's instructions and automated on EpMotion automated liquid handlers (Eppendorf, Germany). Sequencing was performed on the Illumina NovaSeq 6000 sequencing platform with paired-end 150 bp cycles at the Institute for Genomic Medicine (IGM), UC San Diego.

Microbiome data analysis

Demultiplexed fastq files provided by the UC San Diego Microbiome Core and sample metadata are available on the QIITA⁷⁴ open-source microbiome study management platform under the study ID 15161 and ENA study project PRJEB76409 (accession ERP160935). Files were processed using the default workflow, involving trimming of the autodetected adapters, filtering of reads mapping to the mouse genome, and the generation of the operational genomic unit (OGU) and KEGG orthologous tables and using Woltka.⁷⁵ R version 4.2.2 (R Foundation for Statistical Computing, Vienna, Austria) was used for the downstream analysis. Data was manipulated using “phyloseq v 1.42.0” and OGUs were collapsed at species level. Data was robust centered log ratio transformed using the decostand function from “vegan v 2.6-4” and taxa with zero variance between samples were removed using the nearZeroVar function from “caret v 6.0-94.” PCA was performed using “mixOmics v 6.22.0” and centroid separation was evaluated using PERMANOVA from the adonis2 function from “vegan v 2.6-4.” Group dispersion was checked for satisfying PERMANOVA assumptions using the function betadisper from “vegan v 2.6-4.” PLS-DA was performed using “mixOmics v 6.22.0,” and models were evaluated using leave-one-out cross-validation to calculate the CER. Finally, pairwise differential abundance analysis for taxa and enzymes was performed using ALDEx2 from “CoDaSeq v 0.99.6.” Taxa with P-value <0.05 were retained for interpretation.⁷⁶

Author contributions

A.T.A., M.R., and P.C.D. conceptualized the idea. A.T.A. performed LC-MS/MS analysis. A.K. performed analysis of metabolomics data; S.Z. performed analysis of whole genome sequencing data. K.M. and A.K. performed statistical analysis. H.Z. performed all animal work, and M.R. supervised animal work. A.K. and A.T.A. wrote the manuscript. All authors contributed to editing the manuscript.

Conflicts of interests

P.C.D. is an advisor and holds equity in Cybele and Sirenas and a Scientific co-founder, advisor and holds equity to Ometa, Enveda, and Arome with prior approval by UC-San Diego. PCD also consulted for DSM Animal Health in 2023.

Data availability

All untargeted LC-MS/MS data used in this study are publicly available at MassIVE (https://massive.ucsd.edu/) under the following accession numbers: MSV000084783 (doi:10.25345/C51M4J). The feature-based molecular networking job is publicly available at GNPS: https://gnps.ucsd.edu/ProteoSAFe/status.jsp?task=d35b583c2cf3435289a0aca8de4df4dd. All codes used for analysis are found in the Github repository: https://github.com/k-anastasiia/Iron_diet.