Effect of a complementary feeding intervention based on iron- and zinc-biofortified pearl millet on the gut microbiota in 12–18-month-old children: a randomized trial Authors: Samantha L Huey, Nathaniel L Cole, Ioanna Pagani, Antonio González, Julia L Finkelstein, Jere D Haas, Shobha A Udipi, Padmini Ghugre, Ramesh D Potdar, Rob Knight, Saurabh Mehta Published: 2026 ABSTRACT AbstractIron supplementation studies in children under five suggest potentially adverse gut microbiota shifts. Given iron’s importance during early childhood, food-based approaches may offer a more viable supplementation strategy. Using shotgun metagenomics, we examined the effects of 9 months’ daily consumption of iron/zinc-biofortified pearl millet (FeZnPM; 8.70 mg/100 g iron) versus control pearl millet (CPM; 3 mg/100 g iron) on the gut microbiome in 12–18-month-old children without severe anemia (hemoglobin ≥9.0 g/dL) from Mumbai urban slums through a randomized controlled trial, the primary outcomes of which were iron status biomarkers and infant growth (ClinicalTrials.gov ID: NCT02233764). In paired (n = 81) and endpoint (n = 108) analyses, FeZnPM consumption was not associated with detectable adverse effects on developing microbiomes and exploratory analyses suggest that it may support beneficial metabolic adaptations via direct modulation of antibiotic, antioxidant, and pollutant degradation pathways. This suggests biofortified crops could provide a sustainable approach to addressing iron deficiency while maintaining healthy microbiome development in early life.Subject terms: Microbiome, Paediatric research Introduction Children need adequate iron early in life, particularly after neonatal stores are depleted at around six months of age, in order to grow, mature, and develop1,2. Iron deficiency leads to anemia, poor motor and cognitive development and performance, impaired immune function, and lower work performance. In many contexts where iron deficiency and anemia are common, improving iron status through supplements and fortificants require consuming large doses (e.g., 50–62.5 mg FeSO4) to deliver the recommended 10–12.5 mg elemental iron for preventing iron deficiency and anemia in this age group3. About 80-90% of enteral nonheme iron is not absorbed4–6 and transmits to the colon, where it can serve as an essential nutrient for bacteria, including pathogens such as Escherichia coli, Salmonella, and Shigella7. In addition to iron-induced redox cycling in the gut lumen and mucosa8, greater abundances of these taxa may promote inflammation and contribute to host susceptibility to adverse gastrointestinal symptoms caused by high dose iron supplementation, such as diarrhea, nausea, constipation, abdominal pain, and black or tarry stools in both children and adults9,10. The adverse gastrointestinal effects of unabsorbed iron thus pose a challenge for classic modes of iron supplementation and fortification11,12 to address iron deficiency anemia, compared to the lower amounts of iron found in unmodified foods consumed in typical diets. Further complicating this is that studies (Supplementary Data 1;11–21) have reported microbiota shifts towards pathogenic and pro-inflammatory bacteria after iron supplementation, particularly when sanitation is poor and pathogen exposure is already high11–14,22. Given the scale of these fortification and supplementation programs, and the number of children they affect23, it is critical to examine if other approaches to deliver additional iron have similar adverse effects. One such approach to increase iron intake is through biofortification of staple crops: crops that have been traditionally bred, agronomically fertilized, or genetically engineered to have higher micronutrient content compared to conventional crop counterparts24. Food-based iron from biofortified beans, pearl millet, and rice has been shown to improve iron status across children, adolescents, and adults25–27. However, the effect of iron-biofortified crops on the gut microbiome has not been previously reported in humans. Additionally, there are several more unknowns in microbiome literature, particularly among children. Most gut microbiome studies in children have examined the neonatal period and/or the first year of life, finding that the gut microbiome stabilizes towards an adult-like composition at around age 2 or 328–35, despite few studies reporting on the microbiome dynamics after 12 months of age36. From these studies, the taxonomic composition of the gut microbiome transitions from a high abundance of human milk oligosaccharide-metabolizing bacteria such as Bifidobacteria spp. to a greater diversity of taxa37. This is in response to the cessation of breastmilk and introduction of solid food during infancy, typically around 4-6 months of age38,39, which is affected by a multitude of factors, including delivery mode, exposure to antibiotics, diet, and maternal health40. However, most of these studies were done in high-income settings and therefore may not be generalizable to lower-resource settings, such as urban slums, where there is a higher likelihood of poor nutrition and sanitation, and greater inflammation resulting in gut microbial immaturity and delayed microbial development41. The gut microbiome in lower-to-middle-income country settings remains understudied; a recent article cited that 70% of published human microbiome data comes from <15% of the global population, namely Europe and North America42. Fewer still are longitudinal studies done in children, and randomized controlled trials (RCTs) of dietary interventions are especially scarce36. Given the importance of iron during early childhood, it is essential to test whether food-based approaches, such as biofortification, can be a viable strategy to deliver this nutrient without negative effects on the gut microbiome, and to appropriately inform public policy and programs. Our study aimed to determine (1) the effects of 9 months’ daily consumption of iron-biofortified pearl millet (FeZnPM)-based complementary foods compared to conventional foods on the gut microbiome in children, and (2) characterize the development of the gut microbiome during the 2nd year of life, as a secondary analysis of a RCT43. We hypothesized that (1) provision of FeZnPM-based foods is a non-pharmacological method to improve iron status without causing adverse effects on the gut microbiome and (2) similar to our baseline preliminary data based on 16S rRNA gene amplicon sequencing44, children will have high abundances of species from the Pseudomonadota phylum (formerly Proteobacteria45) at baseline and will undergo shifts in species composition, diversity, and functional capacity at future timepoints. In this study, we show that the consumption of iron-biofortified pearl millet among children in the 2nd year of life is not associated with detectable adverse effects on the developing gut microbiome, compared to conventional pearl millet (CPM). Additionally, biofortified pearl millet may also support beneficial metabolic adaptations compared to control pearl millet from exploratory analyses. Our findings suggest that iron-biofortified crops may be a sustainable approach to address iron deficiency while maintaining healthy microbiome development in early life. Results Baseline gut microbiome in the 2nd year of life Demographic and clinical characteristics for 168 children with available baseline stool samples are reported in Supplementary Data 2. Antibiotic use and history of illness were not systematically assessed (Supplementary Data 3 and 4). However, there were no associations found between antibiotic use and a higher frequency of illness (Chi squared (χ2) and Cramer’s V association tests) (Supplementary Data 5–19). Baseline microbiome composition and α- and β-diversity characteristics, overall and by intervention arm, are shown in Supplementary Data 20. Although β-diversity indices had adjusted p-values < 0.001, the magnitude of these differences were deemed negligible using the Vargha and Delaney A (VDA) effect size measure (VDA > 0.4) (Supplementary Data 21), and therefore not considered biologically meaningful. There were no differences in baseline α-diversity between the n = 15 children who were reported to have used antibiotics and the n = 10 who were reported to have not used antibiotics (Supplementary Data 22). KEGG pathways were filtered to contain at least 70% of their genes (70% pathway completeness) to increase the likelihood of these pathways being expressed by the microbiota. Across baseline, 84 KEGG functional pathways with 70% pathway completeness, and 74 generic GO terms (molecular function and biological processes only) were found. Individual relative abundances of KEGG pathways are shown in Supplementary Fig. 1. The percentage of individuals with particular KEGG pathway or GO terms are shown in Supplementary Figs. 2 and 3. Impact of intervention on the gut microbiome A total of 399 stool samples (N = 199; 1–3 sampling timepoints per child) from both intervention arms across 9-months follow-up were available for microbiome sequencing. No effect of iron-biofortified intervention on α- and β-diversity Relative abundances and α- and β-diversity did not differ between intervention arms at either time point (Supplementary Data 23). Although UniFrac distances between arms were significantly different at an adjusted p-value < 0.0001, the effect size of this difference was deemed as negligible (VDA > 0.4) (Supplementary Data 24), and therefore not considered biologically meaningful. Differences by arm in α-diversity were investigated using random intercept models and estimated marginal means (EMMs). There was no significant difference between intervention arms (at either timepoint), nor did participant age at baseline contribute significantly to the model. Time was the only coefficient that significantly impacted changes in α-diversity (p < 0.05) (Fig. 1 and Supplementary Data 25–28), which increased at a rate that was not affected by intervention arm (determined using difference in difference analysis (Supplementary Data 29)). Microbiota α-diversity metrics by intervention arm and over time. α-diversity was investigated using random intercepts models with intervention arm, time, age at baseline, and the interaction of intervention × time as fixed effects. Data represent n = 81 biological replicates (independent subjects) with 162 total observations (81 subjects × 2 timepoints). Individual regression lines are plotted and colored by intervention arm (conventional pearl millet (CPM) = red; iron-zinc biofortified pearl millet (FeZnPM) = blue). The black dotted line represents the overall regression line. The emmeans package was used to determine if α-diversity estimated marginal means changed significantly over time or between intervention arms. Statistical testing used two-sided tests with Sidak adjustment for multiple comparisons within each model. The p-values depicted on the graph are those of the emmeans contrast testing (either endpoint vs baseline or FeZnPM vs CPM). There were no differences between intervention arm, age at baseline, or the interaction of intervention arm × time (p > 0.05) for any α-diversity metric. All α-diversity metrics increased over time, with estimated marginal means being significantly higher at endpoint compared to baseline (p < 0.001): A Chao Abundance Index: t-statistic[degrees of freedom = 79] = 2.5, p = 0.015, 95% confidence interval = [0.03, 0.23], partial eta-squared effect size = 0.14; B Shannon diversity index (SDI): t-statistic[degrees of freedom = 79] = 3.79, p < 0.001, 95% confidence interval = [2.05, 6.59], partial eta-squared effect size = 0.31; C Pielou’s Evenness Index: t-statistic[degrees of freedom = 79] = 3.47, p < 0.001, 95% confidence interval = [0.03, 0.1], partial eta-squared effect size = 0.28; D Faith’s Phylogenetic Diversity (PD): t-statistic[degrees of freedom = 79] = 4, p < 0.001, 95% confidence interval = [0.09, 0.27], partial eta-squared effect size = 0.33 (Supplementary Data 25–28). Iron biofortified intervention modified OGU abundances We investigated the longitudinal impact of intervention on OGU abundance using zero-inflated generalized linear mixed models (ZI-GLMMs) and subsequent EMM contrast testing. A total of 67 OGUs showed significant differences in abundance between intervention arms with an effect size of medium or more (Fig. 2A). Only 2 OGUs show significant differences in presence/absence patterns (modeled by zero-inflation component) (Fig. 2B). Comparison of bacterial operational genomic unit (OGU) abundance changes between intervention groups. Zero-inflated models were used to estimate OGU abundance changes over time in both intervention groups using two-sided statistical tests with Benjamini–Hochberg correction for multiple comparisons. Data represent n = 81 biological replicates (independent subjects) with zero-inflated generalised linear mixed models fitted to n = 318 OGUs. Data are presented as estimated marginal mean values ± standard error (SE). For brevity, only OGUs with significant differences (FDR-corrected q ≤ 0.05) and Cohen’s d > 0.5 (medium or greater effect size) are shown. A Estimated marginal mean OGU abundance differences at endpoint in the biofortified pearl millet (FeZnPM) intervention arm compared to conventional pearl millet (CPM). B Zero-inflation model component showing significant differences (FDR-corrected q ≤ 0.05) with Cohen’s d > 0.5. The zero-inflation component models the probability of excess zeros, helping to determine structural or random zeros. Data are color-coded based on the phylum the species is a member of: Actinobacteriota = light blue; Bacillota = light red; Bacteroidota = green; Desulfobacterota = dark green; Myxococcota = light red; Pseudomonadota = dark red. Longitudinal abundance changes due to intervention were accessed using the ANCOM-BC method. This identified 13 significantly decreased and five significantly increased taxa in the FeZnPM arm compared to CPM (false discovery rate (FDR) ≤ 0.05; fold change (FC) ≥ 2) (Fig. 3). None of these OGUs were found to be common between the two methods. Cross-sectional differential abundance at endpoint (n = 108), identified 47 decreased taxa and 8 increased taxa with a FC ≥ 1.2, and 17 decreased OGUs and three increased taxa with a FC ≥ 2 between intervention arms (Supplementary Figs. 4 and 5). Bacterial operational genomic unit (OGU) differential abundance changes over time associated with intervention. Data represent n = 81 biological replicates (independent subjects) analyzed using ANCOM-BC differential abundance analysis with two-sided statistical tests and built-in false discovery rate correction. Significant OGUs with a fold change ≥ 2 and FDR-corrected q values ≤ 0.05 in the iron-zinc biofortified pearl millet (FeZnPM) intervention arm at endpoint compared to conventional pearl millet (CPM) are plotted as the natural log of fold changes. Bars are color-coded based on the phylum the species is a member of: Actinobacteriota = blue; Bacillota = red. Impact of intervention on microbiota functional pathways Little difference was found between the number and type of unique pathways in each intervention arm at both timepoints (KEGG pathways: Baseline: CPM = 80 and FeZnPM = 80; Endpoint: CPM = 81 and FeZnPM = 79). For brevity, KEGG pathways present in less than 95% of samples at baseline and endpoint are shown in Supplementary Fig. 6. In line with our microbiota composition results (Supplementary Data 23), there were also no global KEGG pathway or GO term abundance differences between intervention arms at baseline (PERMANOVA KEGG: R2 = 0.01, F-statistic = 0.781, p-value = 0.517; PERMANOVA GO: R2 = 0.016, F-statistic = 1.27, p-value = 0.244), nor at endpoint (n = 81) (PERMANOVA KEGG: R2 = 0.008, F-statistic = 0.616, p-value = 0.666; PERMANOVA GO R2 = 0.005, F-statistic = 0.361, p-value = 0.968). To further interrogate intervention effects on functional capacity, we compared FeZnPM and CPM at endpoint using differential abundance testing, network analysis, and Pearson correlation coefficients for all pathways with ≥ 70% gene coverage. This revealed distinct microbial metabolic adaptations linked to higher iron and zinc intake from biofortified crops. Differential abundance testing showed significant increases (FDR ≤ 0.05; FC ≥ 2) in polyketide backbone and sugar unit biosynthesis and caffeine metabolism and significant decreases (FDR ≤ 0.05; FC ≥ 2) in aminobenzoate and naphthalene degradation and prodigiosin, acarbose and validamycin biosynthesis (Fig. 4). Functional pathway changes over time associated with intervention. Data represent n = 81 biological replicates (independent subjects) with KEGG pathways analyzed using limma/edgeR with two-sided tests and Benjamini–Hochberg p-value correction for multiple comparisons. Barplot shows significant (fold change ≥ 2 and FDR-corrected q values ≤ 0.05) KEGG pathway differences in the iron-zinc biofortified pearl millet (FeZnPM) intervention arm at endpoint compared to conventional pearl millet (CPM). Data are shown as log2(fold changes). Pathway terms are colour-coded according to their parent functional class: Metabolism of terpenoids and polyketides = orange; Biosynthesis of other secondary metabolites = green; Xenobiotic biodegradation and metabolism = beige. Comparing microbial functional pathways between the endpoint FeZnPM and CPM cohorts revealed microbial metabolic adaptations (Fig. 5C and D) in exploratory analyses. Increased pathways in the FeZnPM cohort included (a) biosynthesis of secondary metabolites and antibiotics such as type II polyketide backbone, polyketide sugar, monobactam, and phenazine; (b) biosynthesis of antioxidants such as carotenoid and glutathione, or redox-sensitive signaling and immunologically active molecules such as lipoic acid and O-antigen nucleotide sugar; (c) tryptophan metabolism; (d) benzoate degradation; (e) C-5 branched dibasic acid metabolism; and (f) nitrogen metabolism. Conversely, decreased pathways in the FeZnPM cohort included (a) xylene degradation and aminobenzoate degradation, which are involved in the breakdown of aromatic compounds; (b) mannose-type O-glycan biosynthesis, which contributes to microbial cell wall synthesis; and (c) D-amino acid metabolism, which plays a role in microbial signaling and cell wall components. Differential network analysis of microbial metabolic pathways. Hierarchical clustering heatmap of differentially correlated KEGG pathways for A Endpoint (n = 108) vs. Baseline (n = 167) (iron-zinc biofortified pearl millet (FeZnPM) and conventional pearl millet (CPM) combined) and C Endpoint FeZnPM (n = 56) vs. CPM (n = 52) cohorts. The heatmaps display Pearson correlation coefficient differences, with clustering performed using the Ward method. Color intensity represents the magnitude of correlation differences (red: positive, blue: negative, vmin = −1, vmax = 1).. Differential network visualization for B Endpoint vs. Baseline and D Endpoint FeZnPM vs. CPM, generated using Gephi with the Yifan-Hu layout algorithm. Node size is proportional to weighted degree centrality, reflecting pathway connectivity, and node color indicates modularity class, highlighting community structures within the network. Edges are colored on a blue-to-red scale based on the magnitude of positive Pearson correlation coefficient differences, with blue indicating negative differences and red indicating positive differences (|difference| > 0.2). Iron deficiency, anemia and sex do not affect changes in α-diversity We also investigated the effect of iron deficiency, and/or anemia, on α-diversity using random intercept models, with sex as a covariate. Serum ferritin and hemoglobin measurements were available for 77 paired samples. Like previous results (n = 81 paired samples), changes in α-diversity did not differ between intervention arms (Supplementary Data 30). There were also no cross-sectional differences, nor differences between the rates of change over time among males and females, nor within any particular iron status group, regardless of intervention arm (Supplementary Data 31–33). However, we did observe that SDI and evenness differed at baseline and endpoint between anemia severity groups, but only in males with iron deficiency (Supplementary Data 34), the effect of which appears to increase with worsening anemia severity (Supplementary Fig. 7). Regardless of subgroup, α-diversity statistics significantly increased over time (q < 0.05; Supplementary Fig. 8 and Supplementary Data 35). Gut microbiome development in the 2nd year of life Considering the few differences detected between intervention arms, we investigated the development of the microbiota over time without this grouping to further understand community structures during the 2nd year of life. We hypothesized that, similar to our baseline preliminary data based on 16S rRNA gene Illumina sequencing44, children will have high baseline abundances of Pseudomonadota species (formally Proteobacteria45) and will undergo shifts in species composition, diversity, and functional capacity at future timepoints. We observed a high relative abundance of Pseudomonadota at baseline (a median of 57% relative abundance), which decreased to ~52% (p-value = 0.1, VDA = 0.577 (small effect size); Supplementary Data 21 and 24). In examining changes in abundance of taxa at a prevalence of 40% or more (n = 1990) with ZI-GLMMs, a total of 337 OGU models were successfully fit with either a negative binomial or Poisson distribution modeling the continuous model component (Supplementary Data 36). EMM differences at endpoint compared to baseline (ignoring intervention arm) are shown in Supplementary Data 37. Of these, 179 had an FDR-adjusted p-value ≤ 0.05, 139 of which had a medium or greater effect size (Fig. 6). Change in bacterial operational genomic unit (OGU) abundance over time. Data represent n = 81 biological replicates (independent subjects) analysed using zero-inflated generalised linear mixed models with two-sided statistical tests and Benjamini–Hochberg correction for multiple comparisons. Data are presented as estimated marginal mean values ± standard error (SE). 139 OGUs had a significant change (false discovery rate (FDR) corrected q-value ≤ 0.05) in estimated marginal mean abundance at endpoint compared to baseline with Cohen’s d ≥ 0.5 (medium or greater effect size). Significance levels are indicated by: * q ≤ 0.05; ** q ≤ 0.01; *** q ≤ 0.001. Exact FDR-corrected q-values are provided in Supplementary Data 37. Data are color-coded based on the phylum the species is a member of: Actinobacteriota = light blue; Bacillota = light red; Bacteroidota = green; Desulfobacterota = dark green; Myxococcota = light red; Pseudomonadota = dark red. Differential abundance analysis revealed 1387 OGUs with a significant FC ≥ 1.2 (FDR ≤ 0.05) in abundance at endpoint compared to baseline (Supplementary Data 38). Of these, 627 decreased, and 760 increased; the remaining 847 OGUs did not show any significant changes in abundance. However, to concentrate on the most biologically meaningful changes, we focused on significant FCs ≥ 2 (FC2 & FDR < 0.05). For significant FCs ≥ 2, there were 671 unique species and 335 unique genera (represented by 760 unique OGUs; Supplementary Data 38) that changed in abundance over time. Among the significant taxa identified in ZI-GLMMs EMM testing (Supplementary Data 37) and those with significant FCs ≥ 2 (Supplementary Data 38), 25 OGUs were common between the two analyses, most of which were members of the Bacillota phylum and had increased abundance at endpoint compared to baseline (Fig. 7). None of these taxa overlap with results presented in Fig. 2, indicating that time, and not intervention arm, influences this change in abundance. Overlapping differential abundance changes over time between statistical methods. Data represent n = 81 biological replicates (independent subjects) showing operational genomic units (OGUs) identified as significantly different (FDR-corrected q ≤ 0.05) by both zero-inflated generalized linear mixed models (ZI-GLMM) estimated marginal means and ANCOM-BC differential abundance analysis. Only OGUs with fold change ≥ 2.0 in both analyses are displayed. Both methods used two-sided statistical tests with appropriate multiple comparison corrections. Bars are color-coded based on the phylum the species is a member of: Actinobacteriota = blue; Bacillota = red; Bacteroidota = green; Pseudomonadota = orange. Development of gut microbiota functional pathways over time Initial investigations into the functional capacity of the microbiota involved KEGG annotation and differential network analysis. Differential abundance analysis revealed 6 decreased KEGG pathways at endpoint compared to baseline (aromatic compound, aminobenzoate and naphthalene degradation, carotenoid biosynthesis, and caffeine and sphingolipid metabolism), while secondary metabolite biosynthesis increased (polyketide sugar unit, acarbose and validamycin biosynthesis; Fig. 8). Changes in differential abundance of KEGG pathways over time. Data represent n = 81 biological replicates (independent subjects) with KEGG pathways filtered for ≥70% gene completeness and analyzed using limma/edgeR with two-sided tests and Benjamini–Hochberg p-value correction for multiple comparisons. Eight KEGG pathways were significantly differentially abundant at endpoint compared to baseline (fold change ≥ 2 and FDR-corrected q values ≤ 0.05). Pathway terms are color-coded according to their parent functional class: Degradation of aromatic compounds = purple; Lipid metabolism = blue; Metabolism of terpenoids and polyketides = orange; Biosynthesis of other secondary metabolites = green; Xenobiotic biodegradation and metabolism = beige. Differential network analysis comparing endpoint versus baseline in FeZnPM and CPM arms (combined) revealed significant shifts in microbial metabolic pathways (Fig. 5A and B). Among the increased differential pathways were (a) amino acid metabolism such as tryptophan, valine, leucine, isoleucine, and arginine metabolism; (b) carbohydrate metabolism pathways, such as starch, sucrose, and pentose phosphate pathways; (c) lipid metabolism, such as glycerolipid and fatty acid degradation; (d) secondary metabolite biosynthesis, such as carotenoid and folate production, phenazine, streptomycin, monobactam, and type II polyketide backbone biosynthesis; (e) xenobiotic degradation pathways such as aminobenzoate and xylene; and (f) energy and cofactor metabolism such as sulfur and nitrogen. Conversely, decreased pathways included caffeine metabolism, naphthalene degradation, and glycosaminoglycan degradation. Discussion In this study, we characterized gut microbiome differences between children receiving either FeZnPM or CPM during a parent-randomized feeding trial43,46, and examined gut microbiota development during the second year of life. We demonstrated that providing additional iron and zinc via FeZnPM does not lead to deleterious microbiome shifts compared to CPM, while improving hemoglobin concentrations (shown previously43). While provision of iron and zinc is essential—with host zinc status47 and consumption48 associated with benefits to gastrointestinal function and gut microbiota composition—concerns have been raised about the effect of iron supplementation or fortification on the gut microbiota (Supplementary Data 1)11–13,49. Considering its essential role in health, including early-life immune system development, the lack of iron during this crucial stage of child development can have long-lasting adverse effects. This creates a critical dilemma: children need iron, but some delivery methods may have unintended consequences during an important developmental period. We show in this paper that provision of iron and zinc through a food-based approach, while still contributing to iron requirements, is not associated with potentially adverse effects on the gut microbiota. This approach has the added benefit of promoting healthy diets, avoiding supraphysiological or pharmacological doses of iron, and helping to support food systems already challenged by climate change50,51. Based on animal studies, such food-based iron interventions may also positively impact the microbiota and foster the development of a robust microbiota in children52. Little is known about the development of the gut microbiota in children older than 12 months, especially those living in low-resource settings with high burdens of disease and malnutrition. Hence, we aimed to investigate the microbiome at baseline prior to intervention and its development over time. All participants in this study were at least 12 months old at baseline (median = 14.62 months (maximum 19 months); Supplementary Data 2). Despite the high prevalence of undernourishment, Shannon Diversity Index (SDI) (median = 3.34) was within ranges reported in apparently healthy children (Supplementary Data 20 and 23)53–56. Also in line with previous studies53–55,57, we observed with random intercept models that diversity increased with age (p-value < 0.001; Fig. 1), reflecting continued microbiota maturation with improved evenness and reduced dominance by Escherichia and Enterococcus (from ~45 to ~36%). Importantly, diversity trends were consistent between FeZnPM and CPM arms, suggesting that FeZnPM does not disrupt normal gut maturation (increasing diversity) the way traditional iron supplements can (Supplementary Data 1). However, there was a noticeably high level of variance between individuals in both arms. We hypothesized that some of this unexplained variance may be attributable to differences in nutritional status. As the focus of this study was on the impact of consuming an iron biofortified crop, we aimed to investigate any differences in microbiota diversity due to baseline serum ferritin and/or hemoglobin concentration. While few differences in α- or β-diversity emerged between these sub-arms, we observed significant differences in SDI and evenness at baseline and endpoint between anemia severity groups, but only in males with iron deficiency (Supplementary Data 34 and Supplementary Fig. 7). It is likely that there are further differences attributable to other characteristics of this cohort, e.g., antibiotic use, which we were unable to access due to missing data. Additionally, we investigated changes in OGU abundance between intervention arms cross-sectionally at baseline and endpoint, and over time in paired analyses using two approaches: ZI-GLMMs and differential abundance testing with ANCOM-BC. While overall diversity patterns were similar between intervention arms, detailed compositional and functional analysis revealed that FeZnPM consumption influenced the microbiome in ways that may have additional benefits for child health. Most notably, both arms exhibited an increase in Bacillota, which represents a critical transition in early life microbiome establishment56,58. Many members of this phylum play an integral role in metabolic and immune system programming, colonization resistance, complex carbohydrate fermentation and subsequent short-chain fatty acid production (especially butyrate)59, including some of the taxa found to increase in this cohort, e.g., Butyrivibrio, Anaerotruncus, and Blautia60,61. In addition, several taxa may be specifically targeted by FeZnPM compared to control, notably the reduction of pathogenic species from the Campylobacter genus (Supplementary Fig. 5B), Listeria monocytogenes (Fig. 3), and Ralstonia mannitolilytica (Fig. 2A and B), a finding with important implications for children in high-pathogen environments. These potentially beneficial compositional changes contrast with traditional iron supplementation studies, which consistently show increased enteropathogens and decreased beneficial taxa such as Bifidobacterium, Lactobacillus, and Roseburia, particularly in settings with poor sanitation (Supplementary Data 1). Our study cohort represents a high-risk population where iron-driven microbiome disruptions could be amplified, yet FeZnPM consumption demonstrated no such affect compared to CPM. This has significant implications for national nutrition programs: biofortified crops could enable large-scale iron interventions in vulnerable populations without deleterious effects on the microbiome. To understand the functional significance of these compositional changes, we investigated differences in metabolic potential between FeZnPM and CPM arms, as well as overtime. While taxonomic shifts provide important insights, the functional capacity of the microbiome, which determines its actual impact on host health, cannot be inferred from composition alone. Exploratory differential co-occurrence network analysis comparing endpoint and baseline showed a maturing gut microbiome, characterized by enhanced metabolic capabilities and reduced reliance on host-derived substrates, consistent with dietary diversification and microbial succession in both the FeZnPM and CPM arms. Increased pathways such as amino acid (e.g., tryptophan, valine, leucine, isoleucine, and arginine), carbohydrate (e.g., starch, sucrose, and pentose phosphate), and lipid (e.g., glycerolipid and fatty acid degradation) metabolism, alongside biosynthesis (e.g., carotenoid and folate production, phenazine, streptomycin, monobactam, and type II polyketide backbone) and xenobiotic degradation (e.g., aminobenzoate and xylene) (Figs. 8 and 5A and B) indicate a microbiome adapting to diverse dietary inputs and environmental exposures, supporting host physiology through secondary metabolite production. We also compared KEGG metabolic pathways between arms in terms of differential abundances, correlation and co-occurrence at the endpoint. At the endpoint, the FeZnPM arm exhibited differentially increased pathways for antibiotic biosynthesis (phenazine, monobactam) and antioxidant metabolism (glutathione, lipoic acid) (Fig. 5C and D), suggesting enhanced colonization resistance and gut health. Conversely, decreased pathways, including pollutant degradation, such as xylene and aminobenzoate, prodigiosin and mannose-type O-glycan biosynthesis (Fig. 5C and D), align with reduced pathogen presence, such as Campylobacter at endpoint in the FeZnPM arm, as supported by our differential microbial species analysis (Supplementary Fig. 5B and Fig. 2) and reduced Pseudomonadota (Supplementary Data 20 and 23). The reduction of prodigiosin biosynthesis may also suggest a decrease in the opportunistic pathogen Serratia marcescens and Vibrio species, which are known producers62 (Fig. 4). Some observed pathway differences between FeZnPM and CPM require nuanced interpretation, for example, enrichment of caffeine metabolism and reduction of naphthalene degradation. KEGG pathway assignments based on metagenomic data may be subject to annotation ambiguity due to two main factors: (1) incorrect gene assignment from sequence similarity, and (2) mapping of single multifunctional enzymes to multiple metabolic pathways, which could explain the presence of statistically differentially abundant caffeine metabolism and naphthalene degradation in the FeZnPM intervention arm at endpoint. Our KEGG pathway assignments were generated using Woltka63, and to minimise artefactual reporting from partial pathway fragments, we filtered results to retain only KEGG pathways with ≥70% gene completeness. This makes it less likely that differentially abundant results are derived from sparse, incomplete pathways. Nevertheless, some KEGG pathways share substantial enzyme overlap, and broad-specificity enzyme families (e.g., oxidoreductases, hydrolases) can contribute to apparent completeness across multiple pathways64. Enrichment of caffeine metabolism in this cohort is unlikely, however this could reflect tea consumption, which increased from 29% at baseline to 57% at endpoint, or passive caffeine intake through breastmilk65,66. Alternatively, this signal may reflect changes in the abundance of enzymes involved in purine/xanthine oxidation also overlap with caffeine metabolism. Such enzymes can be modulated by other diet-derived metabolites, including aromatic polyphenols such as gallic acid and caffeic acid present in FeZnPM67, underscoring that pathway labels may capture broader metabolic activity rather than specific substrate turnover. Naphthalene is a polycyclic aromatic hydrocarbon used as a pesticide and in numerous industrial applications, and is not a natural metabolite of the gut microbiota. It is primarily excreted via urine, and unreported as a fecal metabolite68,69, thus it is unusual to see naphthalene degradation decreased in the FeZnPM arm (Fig. 5). This may instead reflect concurrent downregulation of aromatic ring–cleaving enzymes or dioxygenases that function in multiple xenobiotic and natural aromatic compound degradation pathways, including polyphenol-derived substrates. In contrast, the observed changes in benzoate degradation (increased) and aminobenzoate degradation (decreased) (Figs. 4 and 5C and D) are biologically plausible given the dietary context. The Dhanashakti pearl millet variety developed by HarvestPlus and used in this study contains significantly more total polyphenols than its conventional counterpart67,70, which may partially explain the increase in benzoate degradation (Figs. 4 and 5C and D). Benzoates are also common byproducts or intermediates of gut microbial polyphenol metabolism71,72, and degradation pathways serve to detoxify these compounds, mitigate their antimicrobial effects, or channel them into central metabolism via acetyl-CoA64,72. Conversely, decreased aminobenzoate degradation may indicate preferential retention of aminobenzoate derivatives for biosynthetic functions. Aminobenzoates such as anthranilate (derived from indole degradation) and p-aminobenzoic acid (a folate intermediate)73 link these pathways to aromatic amino acid biosynthesis (e.g., tryptophan), folate metabolism, and overall microbial nutrient processing. We observed increased tryptophan metabolism in the FeZnPM group, likely yielding more indole derivatives74, which could be conserved rather than degraded based on the observation of decreased aminobenzoate degradation (Figs. 4 and 5C and D). Such conservation may reflect a functional repurposing of aminobenzoates for folate and amino acid production, microbial signalling molecules, or secondary metabolites, rather than directing them into catabolic routes and supported by increased micronutrient and polyphenol availability from FeZnPM. Notably, these aromatic pathways are also heavily interconnected with fatty acid metabolism and central carbon metabolism (e.g., the TCA cycle)64,72,73, making them sensitive indicators of aromatic compound turnover and metabolic adaptability. These findings demonstrate that micronutrient-biofortified crops-based foods promote a resilient microbiome, with implications for improved immune development, metabolic health, and long-term disease prevention in children. However, as we are unable to validate these metagenomic KEGG annotations in vitro, these interpretations remain speculative. In this RCT, iron-biofortified pearl millet consumption was not associated with detectable adverse effects on the developing gut microbiome among children in the 2nd year of life, and may support beneficial metabolic adaptations compared to control pearl millet lower in iron content. These findings suggest iron-biofortified crops may be a sustainable approach to address iron deficiency while maintaining healthy microbiome development in early life. Further, the child gut microbiome appears to continue developing with changes in diversity, relative abundance, and functional pathways across the 2nd year of life, highlighting the importance of gut microbiome assessment after age 12 months of age as well as potentially new critical windows for interventions during the first 1000 days of life. Limitations Antibiotic use The parent RCT did not comprehensively assess certain variables such as antibiotic use. Given the limited information on antibiotic usage, we cannot exclude the possibility of its influence on our results. Assuming effective randomization for all variables, including antibiotic use, these findings may still have real-world implications where antibiotic use is not controlled. Selection bias The children (n = 199) in this analysis may have different characteristics compared to the children (n = 24) who did not provide stool samples. To investigate this, we compared their baseline characteristics and found no differences in age, sex, or other sociodemographic characteristics (Supplementary Data 2). Generalizability Our results may not apply beyond children eligible for the parent trial (12–18 months of age, living in urban slums with hemoglobin concentration ≥ 9 g/dL. Severely undernourished (weight-for-length Z-score [WLZ] < −3) children, those with hemoglobin concentrations <9 g/dL, and/or who were reported to be taking iron and zinc supplements prior to the start of the trial were excluded, as described previously43. Children with severe or moderate acute malnutrition reportedly have distinct microbiomes compared to healthy children41,75,76. Hence, our results may not extend to severely or moderately undernourished children living in urban slums. However, we anticipate our results to be generalizable to children in the 2nd year of life with either no anemia (hemoglobin concentration >10.5 g/dL77 or mild-moderate anemia (hemoglobin concentration 9–10.5 g/dL), with WLZ > −3 and not consuming iron or zinc supplements, which represented the majority of children screened for the parent trial43. While our work may not be generalizable to children living outside of urban slums, we note that the number of individuals, including children living in urban slums worldwide, is growing. The United Nations estimates that over half the global population lives in urban areas, which is anticipated to increase beyond 65% by 205078,79. In Mumbai, the Indian 2011 census estimated that over 40% of the population resides in urban slums78. In 2020, 117 million children under 5 years of age (49% of the urban population) across India lived in slums80. Therefore, these findings may be generalizable to a large proportion of the world’s children, who are living in urban slums. Dose and bioavailability Dose and bioavailability (the amount of a nutrient absorbed by the body), depends on many factors, including host iron status, inflammation, and anti-nutrients like phytates81. Iron bioavailability in FeZnPM may have contributed to the lack of adverse effects and observed microbiome differences compared to CPM. Each child consumed ~3.0 mg iron along with ~1.6 mg zinc from the intervention daily for 9 months43. This food-based iron dose is smaller than typical for iron interventions. For example, multiple micronutrient powders used in nutritional programs contain 12.5 mg of iron82 which several studies show increased pathogen abundance and intestinal inflammation compared to no-iron controls (Supplementary Data 1), likely due to greater amounts of unabsorbed iron reaching the colon—about 80% or 10 mg13. Biofortification results in greater micronutrient content and similar bioavailability compared to conventional counterparts24. Therefore, iron absorption is improved due to increased iron content when consuming iron from FeZnPM; the availability of unabsorbed iron from nonheme sources like FeZnPM, for colonic gut microbiota to utilize, has been difficult to measure and predict83. In a previous trial among Indian adolescents, 28% of whom had anemia and 41% of whom had iron deficiency, FeZnPM yielded about 7.4% iron absorption, similar to CPM at 7.5%24. In Indian children <2 years of age with iron deficiency, Kodkany and colleagues found that iron absorption from FeZnPM was 3× that of CPM84. While we did not examine iron absorption specifically in the participants of the current study, 65–70% of the population had anemia and/or iron deficiency, suggesting higher iron absorption due to increased needs. Concurrently, any unabsorbed iron did not adversely impact the gut microbiota, a finding in line with animal studies previously described. Experimental validation The results presented in this manuscript are based solely on statistical analyses. Further research using in vitro and in vivo experimentation could elucidate mechanisms of action and provide additional supporting evidence for the FeZnPM-specific changes in composition and metabolic networks. Statistical power This microbiome analysis is an exploratory secondary analysis nested within a larger randomized feeding trial that was powered primarily for nutritional outcomes, rather than microbiome‑specific outcomes. Therefore, our sample size and design constraints stem from the parent RCT, and serves to generate hypotheses, identify potential associations, or characterize patterns rather than to formally test pre‑specified hypotheses with pre‑calculated sample sizes. To assess the sensitivity of our study to detect the observed longitudinal effects in selected variables, we performed a post‑hoc power analysis using the Evident framework85. Power (1–β) was estimated for varying sample sizes and significance thresholds (α = 0.01, 0.05, and 0.1). As illustrated in Supplementary Fig. 9A–F, outcomes showed variable‑specific differences in statistical power. Notably, α-diversity (SDI and evenness) reached or exceeded the conventional ≥ 0.8 power threshold with ≥ 50 observations at α = 0.05. In contrast, taxa abundances (Bifidobacterium and Escherichia) exhibited relatively low power across the studied sample sizes, with maximum values below 0.3 even at α = 0.1, indicating limited sensitivity to detect effects for these endpoints under the observed effect sizes. Any true differences, if present, would require larger sample sizes or stronger effect sizes to be reliably detected. These findings highlight that while diversity metrics, particularly evenness, were well‑powered in this dataset, microbiome taxonomic variables were under‑powered, underscoring the importance of considering variable‑specific detectability when interpreting non‑significant results. Given the low estimated power for certain variables, our future work will include validating the results observed in this exploratory study in a larger cohort. Strengths To our knowledge, this study is the first to report on the impact of consuming biofortified crops on the human gut microbiome, filling in a critical gap in the literature on safety and efficacy of biofortification for addressing micronutrient deficiency. Additionally, it utilizes shotgun metagenomic sequencing, enabling species-level identification and direct functional inference, compared to 16S rRNA gene sequencing86, giving a more informative view of the impact of intervention on the microbiome. Finally, over 80 children had paired microbiome samples available, allowing us to assess longer-term development in the 2nd year of life between 12 and 27 months of age in a relatively large cohort. This is in contrast to a number of nutrition intervention trials in children that ceased follow-up beyond 12 months of age36. Methods Ethics The study protocol was approved by the Intersystem Biomedica Ethics Committee (ISBEC, Mumbai, Maharashtra) (ISBEC/NR-17/ KM-JVJ/2014) and the Institutional Review Board at Cornell University. The study protocol was reviewed and approved by the Health Ministry Screening Committee of India (Indian Council of Medical Research). The parent trial was registered at ClinicalTrials.gov (NCT02233764) and Clinical Trial Registry of India (REF/2014/10/007731). Study design and participants This analysis is nested within a parent randomized controlled parallel arm trial that aimed to address anemia and iron deficiency in children living in urban slum communities of Mumbai (ClinicalTrials.gov ID: NCT02233764)43. The trial, “Effect of Iron/Zinc-Biofortified Pearl Millet on Growth and Immunity in Children Aged 12–18 Months in India” was funded by HarvestPlus/International Food Policy Research Institute and the microbiome analyses was partially also supported by the USDA43. The trial was conducted between March 2017 and July 201843. While examining the impact of intervention on the gut microbiome was not a primary outcome, we planned and collected samples to be able to conduct these analyses in the future46. Eligible children living in urban slums whose caregivers provided informed consent to be screened were enrolled (as described in detail previously43,44,46). Inclusion criteria were children aged 12–18 months and hemoglobin > 9 g/dL. Exclusion criteria were children with an unavailable caregiver, known chronic disease, severe malnutrition defined as weight-for-height z-score <−3, hemoglobinopathies, known dietary allergies, and prior consumption of iron or zinc supplements. For the current study, we also examined information on the number of children who had taken antibiotics (currently or recently) as reported by caregivers, past morbidities such as report of fever, diarrhea, or cough, and birth weight. Participants (n = 223; Supplementary Fig. 10) were randomized to receive either the biofortified intervention or the conventional, non-biofortified crops-based complementary foods, 2 to 3 times daily, 6 days per week, for 9 months (maximum age at endpoint: 27 months)43. The complementary foods were traditional Indian recipes previously standardized for acceptability among mothers and their children in the targeted population87. These foods were prepared with either iron- and zinc-biofortified pearl millet (FeZnPM) or non-biofortified CPM. With a median (IQR) intake of 3.0 (2.7, 3.1) mg iron and 1.6 (1.5, 1.7) mg zinc per child per day43, the FeZnPM intervention delivered ~50% of the estimated average requirement for iron and zinc (6.0 and 2.8 mg/d, respectively)88. Data collection Main follow-up assessments, including 24-h dietary recalls, anthropometric measures, and clinical exams, were done at baseline (pre-intervention), monthly, and endpoint at 9 months. Rectal swabs for microbiome analysis and blood samples for determination of nutritional biomarkers and immune parameters were collected at baseline, endpoint, and at a random midpoint (one of the 7 intervening months of the feeding trial, wherein the participant was serially sampled randomly with no replacement so that each child provided only one more biological sample collection in addition to baseline and endpoint)88,89. Data were collected by using our custom electronic database capture system, ConnEDCt, and stored on a secure server prior to analysis90. Stool sample processing Fecal samples (n = 399 samples across up to 3 timepoints) were collected using rectal swabs [Copan FecalSwab™ Regular Flocked Collection Kit (Nylon FLOQSwab and tube containing 2 mL Cary-Blair medium) (Thermofisher, catalogue #R723487)] by inserting the swab 2 to 3 cm into the rectum and rotating 360 degrees until fecal material was visible on the swab. Two separate swabs per participant were collected to ensure maximal DNA recovery. Samples were transported within 1–6 h after collection to SRL Laboratories in Mumbai, for storage at −80 °C until data collection was completed. Samples were then shipped to Genotypic Technology Pvt. Ltd., Bangalore, for DNA extraction (Modified Qiagen DNeasy Blood & Tissue Kit + RNase treatment), library preparation and sequencing in batch to minimize batch effects. From the rectal swabs, fecal samples (approximately 300 μL) were transferred to sterile microtubes. Lysozyme (200 μL, 10 mg/mL; Sigma# L6876) was added to the samples and incubated at 37 °C for 30 min. AL buffer (200 μL) was added to the samples and vortexed briefly. The samples were then subjected to Proteinase K treatment at 56 °C for 2 h followed by RNase (MP Biomedicals Cat#210107683) treatment, at 65 °C for 20 min. The lysate was mixed well with 95–100% alcohol and loaded onto the Qiagen DNeasy blood and tissue column (#69506). The rest of the protocol was followed as per the manufacturer’s guidelines. DNA was eluted in 50 μL of 10 mM Tris.Cl, pH8.0 (Sigma #T3038). The quantification and quality of the genomic DNA was assessed using Nanodrop2000 (Thermo Scientific, USA), Qubit (Thermo Scientific, USA) and agarose gel electrophoresis. Library preparation was performed using the QIASeq DNA Library Preparation Protocol (Cat#180475) according to the manufacturer’s instructions. Qubit quantified DNA (2–10 ng was enzymatically fragmented, end-repaired and A-tailed in a one-tube reaction using the FX Enzyme Mix provided in the QIASeq FX DNA kit. The end-repaired and adenylated fragments were subjected to adapter ligation, whereby index-incorporated Illumina adapter was ligated, to generate sequencing libraries. These libraries were subjected to 10 cycles of Indexing-PCR (Initial Denaturation at 98 °C for 2 min, cycling (98 °C for 20 s, 60 °C for 30 s, 72 °C for 30 s) and final extension at 72 °C for 1 min) to enrich the adapter-tagged fragments. Finally, the amplified libraries were purified using JetSeq Beads (Bio, # 68031). Further, the Illumina-compatible sequencing libraries were quantified by Qubit fluorometer (Thermo Fisher Scientific, MA, USA). Molar concentrations of the libraries from each sample were normalized and multiplex pools were prepared. The fragment size distribution of the final sequencing library pools was analyzed on Agilent 2200 TapeStation. Barcoded DNA fragment libraries were generated with 0.25 ng of input DNA using a KAPA HyperPlus (Roche) sample preparation kit per manufacturer’s instructions, as the Knight Lab’s internal benchmarks have shown this kit to be better than Nextera for assembly. A BioAnalyzer (Agilent) was used to determine the distribution of fragment sizes (predicted to be between 430 and 990 bp). Samples were pooled with final, single-stranded normalized libraries and sequenced on Illumina MiSeq to obtain the cluster number and stoichiometric distribution of each sample, pooled again and then sequenced using HiSeqX Ten to obtain a set of sequencing reads representing sample-specific metagenomes. In sum, sequences yielded 8 GB data per sample (25–28 M paired-end reads per sample, with 150 × 2 chemistry). Bioinformatics Bioinformatics analysis was conducted on high-performance computing Linux servers, hosted by Cornell University’s Bioinformatics Facility HPC group, BioHPC, in the Cornell Institute of Biotechnology. Sequence quality control and classification All individual samples in which gut microbiome sequencing was attempted yielded successful sequencing results. Raw sequences were trimmed and filtered to remove adapter sequences and low-quality base calls using fastp v0.23.491. Contaminating human DNA was removed using a custom pangenome reference database called PanDora (available at https://github.com/cpnh/PanDora) by kmer matching using BBDuk v39.03. After quality control, reads were aligned to the Web of Life v2 reference phylogeny (https://biocore.github.io/wol/)92 using Bowtie2 v2.4.2-2. Mapped reads were then assigned to operational genomic units (OGUs), KEGG pathways and GO terms using Woltka, as previously described in ref. 63. Misannotations (GO terms only found in Eukaryotes, e.g., muscle systems) and two obsolete terms (GO:0055065 and GO:0007568), and terms related to the Cellular Component ontology were removed from Woltka results. Results were first filtered using the GO slim prokaryote subset (https://geneontology.org/docs/download-ontology/), and the remaining GO terms were manually checked for bacterial entries using AmiGO (https://amigo.geneontology.org/amigo/search/annotation) and UniProt. Microbial diversity analysis Library sizes were normalized using rarefaction to the number of reads in the smallest sample (1,209,312 reads per sample) using Phyloseq v1.46.0 in R v4.3 before conducting diversity analysis. The α- (Chao index, Shannon Diversity Index (SDI), Pielou’s evenness, and Faith’s phylogenetic diversity (PD)) and β- (Unweighted and Weighted UniFrac, Bray–Curtis, and Jaccard distances) diversity indices were calculated in R v4.3 using vegan v2.6.4 and Phyloseq v1.46.0. β-diversity metrics were visualized by metric multidimensional scaling/principal coordinate analysis using the Phyloseq package, using the distance matrices previously mentioned. Permutational multivariate analysis of variance (PERMANOVA) testing was used to determine if there were any differences between intervention arm baseline (n = 168) and between baseline and endpoint (n = 81) samples. As PERMANOVA assumes homogeneity of multivariate dispersions across groups, Permutational Multivariate Analysis of Dispersion (PERMDISP) tests were run using the vegan betadisper() function. This compared within-group spread (either among intervention arm or timepoint) using the average distance between an individual observation to their group centroid to identify the relative spreads of data cloud shapes among groups. Linear mixed models (LMMs), in the form of random intercept models, were used to examine the effect of intervention arm, FeZnPM, compared to CPM, on α-diversity metrics, using R packages nlme v3.1-164 and emmeans v1.10.1. Briefly, continuous numeric variables including the number of days since starting to consume either CPM or FeZnPM (TIME) and child’s age (number of days old), rounded to the nearest whole (AGE) were scaled and assessed for multicollinearity before adding as model terms. TIME and AGE were highly correlated, hence the effect of age on model outcome was represented as the participants age at baseline (AGE_B). Due to non-normally distributed residuals, some outcome variables required transformation before applying the finalized model. After trying multiple combinations, we determined that squaring SDI and evenness metrics, using the natural log of Chao’s index and Faith’s PD, were most suitable for numerical stability and to meet the assumptions of the model (e.g., normally distributed residuals, constant variance or errors, no multicollinearity). Subsequent model residuals were normally distributed. Random intercept models were applied to the data with individual subject as a random effect, intervention arm, TIME and AGE_B as fixed effects. Parameter estimation was performed using the restricted maximum likelihood method. Contrast analysis was performed with emmeans, which calculates EMMs (also known as least-squares means). Categorical variables were weighted according to their distribution to mitigate any effects of imbalanced data. The emmeans package back-transforms data, allowing users to visualize and interpret data on its original scale. It can also help in interpreting interaction terms. The anova() function was used to test the level of contribution model terms had in explaining the variation of the outcome variable. Given the randomized trial design, we used an intention-to-treat approach and thus did not control for confounding variables in the overall model. Subgroup analysis was performed in a similar manner using random intercept or zero-inflated models, and EMMs, to investigate whether iron and anemia status at baseline, and sex, modified the changes in OGUs over time. Two individuals were removed from this subgroup analysis due to lack of serum ferritin and hemoglobin measurements (complete case analysis). We examined the relationship between α-diversity metrics and iron deficiency (serum ferritin <12 µg/L93), and depletion (serum ferritin <25 µg/L94), anemia (hemoglobin <11 g/dL95), and to reflect the latest WHO cutoffs for this age group, <10.5 g/dL77. We did not impute missing data. OGU differential abundance analysis Given the over-dispersed zero-inflated nature of the OGU abundance data, we opted for zero-inflated generalized linear mixed models (ZI-GLMMs) to estimate changes in OGU abundances over time (between baseline and endpoint). The dataset of 11,353 unique OGUs across paired samples (n = 81, baseline and endpoint) was filtered to remove OGUs with a prevalence less than 40% (resulting in 1992 OGUs). This filtering was applied to ensure that there were still sufficient non-zero datapoints to model abundances in a dataset with low sample size. Several distribution families and their link functions (Poisson, simple negative binomial (linear parameterization), complex negative binomial (quadratic parameterization), Gaussian and log normal) were used to model the continuous component of the model (selected based on model diagnostics), with logit distribution used in the zero-inflated binary component. ZI-GLMMs were fit using the glmmTMB package in R (v 1.1.11), and the emmeans package used to compare groups. To further interrogate changes in OGUs between baseline and endpoint, differential abundance analysis was conducted using the analysis of compositions of microbiomes with bias correction (ANCOM-BC) method. The ANCOM-BC v2.8.1 R package was used for this purpose. OGUs with significant changes in abundance were identified based on p-value (adjusted using the Benjamani–Hochberg (BH) procedure for FDR) and a FC threshold ≥ 2, thus avoiding identifying OGUs with negligible FCs that are still considered statistically significant (i.e., not biologically relevant). Metabolic pathway network analysis Pearson correlation coefficients were computed for KEGG pathways with more than 70% gene coverage, with FDR-adjusted p-values applied to identify significant associations (FDR < 0.05). Networks were generated by comparing correlations between endpoint and baseline timepoints within each intervention arm, as well as between the endpoint intervention arm (FeZnPM) and the endpoint control arm (CPM). For each comparison, significant edges from the respective correlation matrices were used to construct undirected graphs with Gephi v0.10.1, where edges represented Pearson coefficients. By plotting the co-occurrence networks of KEGG pathways in both arms, we identified the most central pathways for each arm by sizing the nodes by between centralities. Betweenness centrality measures how often a node appears on the shortest paths between other nodes in the network, indicating its importance as a connector or intermediary between different parts of the network. A node with high betweenness centrality in a metabolic network serves as a critical hub for linking distinct biological functions and ensuring proper network functionality. Its removal would likely cause major disruptions, breaking the connectivity between pathways and impairing essential metabolic functions, with significant biological consequences. Differential edges were identified as those unique to one network or showing a coefficient difference exceeding |0.2| for shared edges. Network statistics, including weighted degree centrality, were calculated for all pathways, and hierarchical clustering (Ward method) was performed on the differential correlation matrix to generate cluster maps using Seaborn. The top 40 most frequent pathways involved in significant differences (Pearson coefficient > 0.2 or <−0.2) were ranked by occurrence in source/target pairs. Node and edge files were exported to Gephi for visualization, incorporating attributes such as degree differences for node size, modularity class for color, and edge categories (positive/negative) for edge color with Yifan-Hu layout algorithm. Exposure and outcome variables Nutritional biomarkers Blood-based biomarkers were collected and measured as described previously43. Serum ferritin was categorized as iron deficiency (serum ferritin <12 µg/L93), depletion (serum ferritin <25 µg/L94) or replete (serum ferritin ≥ 25 µg/L). Anemia was defined as hemoglobin <1195 and <10.5 g/dL77. C-reactive protein (CRP) was categorized as inflammation when >5.0 mg/L. Serum ferritin concentrations were adjusted for inflammation as described previously43 using the Biomarkers Reflecting Inflammation and Nutritional Determinants of Anemia methods96. We performed random imputation of values between 0 and the limit of detection (LOD) for variables with values below the LOD, such as serum ferritin (LOD: <0.5 mg/L) and CRP < 3.0 mg/L. Intervention arm: The intervention arm was coded as a binary variable representing either the FeZnPM arm or CPM arm. Health history Illness history was collected and measured as described previously44. Any known past antibiotic use was recorded, including antibiotic type and reason for their use. The child’s health history data were reported to the study physician by the mother, a physical examination was conducted by the study clinician, and morbidity data were recorded as a report of the child having had any occurrence of diarrhea, fever, or cough within the past month. Sample size The final analytic sample size (Supplementary Fig. 10) was based on the availability of at least one timepoint for rectal swab samples from each randomized child (n = 199), and longitudinal analyses considered children with paired samples at baseline and endpoint (n = 82 pairs). Reasons for loss to follow-up were unrelated to gut microbiome analyses and are described earlier43. Participants who did not provide a fecal sample for this study (n = 24) did not differ to those that did (n = 199; Supplementary Data 39). Statistics Descriptive statistics were expressed as the median plus IQR for continuous variables and as percentages for categorical variables. Two-sided tests were performed for all statistical analyses and results were considered significant at p ≤ 0.05 or q ≤ 0.05 where correction for multiple testing was applied. Unless otherwise stated, the Benjamini–Hochberg method was used to control the FDR. Sample size for the parent trial was based on assumptions made about mean hemoglobin and serum ferritin values within this population (and assuming 40% loss to follow-up). All study personnel were blinded to intervention group allocation until data collection was completed. A single endpoint sample was excluded from analysis due to issues with the raw sequence reads file, making the file unusable in this analysis. No other data was excluded. Statistical analysis was performed in R v4.3 and Python 3.12. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data Transparent Peer Review file