Saccharomycetes and Malassezia fungi associate with early-life gut maturation and allergic disease risk in childhood

Abstract

While early-life gut bacterial microbiota maturation has been well studied and linked to childhood disease, the development of the gut mycobiome remains poorly understood. Few studies have defined fungal succession in infancy, and even fewer have integrated fungal and bacterial maturation, allowing interkingdom analysis within the same individuals. In this study, we analyzed a subset of the CHILD Study Cohort (n = 1409 participants) and generated both ITS2 amplicon and shotgun metagenomic sequencing data from infant stool samples (n = 2256 samples). We hypothesized that the infant mycobiome follows predictable developmental trajectories that influence childhood health outcomes. We found that fungi are reliable biomarkers for gut maturation, with the notable emergence of Saccharomyces and Malassezia as some of the strongest indicators across both fungi and bacteria. Fungal composition was strongly associated with infant age (R = 0.79, p < 0.001) and with the later development of both atopic dermatitis (adj. p = 0.029) and food allergy (adj. p = 0.013). Further, differences in fungal development coincided with changes in key gut immune-modulating metabolites such as butyrate and glycerol, indicating the functional importance of infant gut mycobiome maturation in early-life immune development. Together, these results highlight the early life mycobiome as a potential therapeutic target to mitigate allergic disease development.

Here, the authors show that infant gut fungi are strong biomarkers of gut maturation, with Saccharomyces and Malassezia as key indicators, further associating fungal development with immune-related metabolites and increased risk of atopic dermatitis and food allergy.

Introduction

The gut mycobiome, comprising fungal populations and their genes within the gastrointestinal tract, is increasingly recognized as an integral component of the human microbiota. Early-life fungal communities appear to be influenced by environmental and dietary exposures and may contribute to the broader process of microbiome assembly^1–3. Fungi also interact with host processes and have been linked to a range of health outcomes, including asthma, atopy, and obesity^1,4–7.

Allergic diseases collectively affect hundreds of millions of children worldwide and continue to increase in prevalence⁸. We recently linked the prevalence of multiple allergic diagnoses at age 5 years to a universal pattern of delayed bacterial maturation in the infant gut microbiome, highlighting the critical role that early bacterial succession likely plays in immune education⁹. The mycobiome may also contribute to the relationship between microbiota development and allergic disease risk^1,4–7,10, with recent work by Bernardes et al. showing that antibiotic exposure can promote expansion of Malassezia spp. in infancy and that Malassezia can enhance allergic inflammation in experimental models¹¹.

Beyond chronic diseases such as atopy and asthma, the mycobiome has also been implicated in invasive fungal infections, where cross-kingdom dysbiosis may create conditions that permit fungal overgrowth and translocation. Zhai et al. demonstrated that Candida bloodstream infections were often preceded by intestinal expansion of pathogenic Candida species alongside depletion of bacterial load and diversity, highlighting the importance of profiling fungal and bacterial communities together¹².

In this study, to define the importance of the early-life mycobiome, we analyzed internal transcribed spacer 2 (ITS2) amplicon sequencing data from 2256 longitudinally collected stool samples from 1409 CHILD study participants. We complemented this with paired shotgun metagenomic profiles of the bacterial microbiome and stool metabolomic profiles within the same infants. Our integrative, large-scale approach enabled a comprehensive interkingdom view of gut microbiota development and its relationship to allergic outcomes. Our findings suggest that fungi may serve as informative temporal biomarkers of early-life microbial development and show a notable association between fungal profiles and the clinical outcomes of atopic dermatitis and food allergy at age five years. These results emphasize the importance of including fungal communities in microbiome studies and suggest that strategies to modify the mycobiome have the potential to influence the development of allergic disease.

Results

The dynamics of the early life mycobiome

The mycobiome begins to assemble in early life, shaped by mode of delivery, breastfeeding status, dietary factors, and environmental exposures^6,13. We explored these dynamics within the CHILD cohort by analyzing stool samples with corresponding ITS2 amplicon sequencing data (n = 2256, collected longitudinally at visits scheduled for 3 months and 1 year of age (Fig. 1a, b; Vancouver: n = 530; Edmonton: n = 370; Winnipeg: n = 776; Toronto: 580). In cohort studies, the absolute timing of sample collection varies due to factors like logistical constraints or participant schedules. Specifically, stool samples from diapers were collected at a home visit at around 3 months [mean (SD), 3.8 (1.1) months] and a clinic visit at around 1 year [mean (SD), 12.5 (1.6) months]. This variation in age at collection was leveraged to generate a temporal picture of taxon changes in the early life mycobiome and to define trends in fungal composition during the first year of life.

Fig. 1. Structure and longitudinal trends of the early-life gut mycobiome.

Early-life gut mycobiome development across the first 1.5 years is shown by age in months for a genera and b species relative abundance in the overall CHILD cohort (n = 2256 stool samples) and by site: Vancouver (n = 530), Edmonton (n = 370), Winnipeg (n = 776), Toronto (n = 580). Genera and species with at least 10% prevalence and 2% abundance across all samples are shown; colors indicate fungal taxa. c Alpha (Observed: Wilcoxon rank sum test p-value = 2e–14; Shannon: Wilcoxon rank sum test p-value < 2.2e–16) and d beta diversity (Bray–Curtis; PERMDISP permutation test: F(1,2254) = 207.91, p = 0.001, 999 permutations) of ASV level abundances between 3-month (n = 1003 samples) and 1-year (n = 1253 samples). Blue indicates 3-month samples and red indicates 1-year samples. e Random Forest fungal-derived age of stool samples plotted against chronological age, with a linear regression line of best fit and 95% confidence interval (3-month samples: n = 1003; 1-year samples: n = 1253; Spearman R = 0.79, p < 2.2e − 16). In regression plots, density is represented by (+) red and (−) blue. Boxplots indicate the median (center line), interquartile range (box bounds), and 1.5 × IQR (whiskers). Points represent individual samples.

From our initial descriptive analysis of 2256 longitudinal samples collected from 1409 participants during the first year of life, several notable trends emerged, including a 48-fold average reduction in the relative abundance of Malassezia and a 14-fold average increase in members of the Saccharomycetaceae family (Fig. 1a). At the species level, we observed increasing abundance of Saccharomycetes sp. and decreasing abundance of Malassezia restricta over the first 18 months of life, along with increases in Rhodotorula mucilaginosa and decreases in an unclassified Cladosporium species (Fig. 1b). Together, these findings suggest that early-life fungal community assembly follows a coordinated developmental trajectory rather than a stochastic process. Taking advantage of the CHILD Study’s multi-site design, we observed that these shifts were largely consistent across geographically distinct populations in Canada, including the cities of Vancouver, Edmonton, Winnipeg, and Toronto.

Although, as shown in Fig. 1, collection times deviated from prespecified sampling points, the majority of samples were clustered around the scheduled 3-month and 1-year time points. This clustering allowed samples to be grouped into the intended 3-month and 1-year categories. In doing so, we discovered an increase in fungal observed diversity, a decrease in Shannon diversity, and distinct alterations of beta diversity across the visit-stratified 3-month and 1-year samples (Fig. 1c, d; Observed: Wilcoxon rank sum test with continuity correction, p-value = 2e–14; Shannon: Wilcoxon rank sum test with continuity correction, p-value < 2.2e–16; beta diversity of 3-month samples n = 1003, 1-year samples n = 1253; PERMDISP permutation test: F(1, 2254) = 207.91, p = 0.001, 999 permutations). Bray–Curtis dissimilarity of the gut mycobiome was associated with antibiotic exposure by 1 year of age, presence of an older sibling, season of birth (with significant associations observed between infants born in summer and other months), and breastfeeding status at 6 months (Supplementary Fig. 1). These results are consistent with previous findings on early-life mycobiome development^5,6,13–21, reinforcing the need for continued study of fungal community dynamics during infancy to further understand the impact of early life influences.

We next examined mycobiome maturation by applying a nested cross-validated random forest model to derive a fungal-based age from early-life ITS2 abundance data. This approach parallels our previous method for modeling bacterial age from species-level shotgun metagenomics data⁹. When applied to fungal data, the model revealed a strong correlation between fungal composition and the infant’s chronological age (Fig. 1e; 3-month samples: n = 1003, 1-year samples: n = 1253; p-value < 2.2e−16, Spearman correlation R = 0.79). Overall, the early-life mycobiome appears interlinked with infant early-life development. Thus, we continued our analyses to explore whether this age-mycobiome link may influence health outcomes of CHILD participants.

Associations between early-life mycobiome maturation and allergic outcomes

Next, we further characterized the mycobiome of CHILD participants to examine its relationship with health outcomes. Specifically, we compared fungal-derived age in infancy between children diagnosed with allergic diseases by 5 years of age (i.e., atopic dermatitis, food allergy, allergic rhinitis, and asthma), and a rigorously defined healthy control group with no evidence of allergy at the 1-, 3-, or 5-year study visits, as previously described⁹. Children in the healthy control group had no allergic sensitization or clinical allergic diagnoses up to 5 years of age, as determined by repeated skin prick testing (SPT) and physician assessment. While fungal-derived age did not differ between groups at 3 months (Fig. 2a), we observed significant differences at 1 year between healthy controls (n = 201) and those with atopic dermatitis (n = 169, adj. p = 0.029) or food allergy (n = 62, adj. p = 0.013) at 5 years (Fig. 2b). These findings suggest that delayed fungal maturation by age 1 year may be associated with increased risk of later allergic outcomes.

Fig. 2. Fungal-derived age and its association with allergic phenotypes.

Boxplots compare fungal-derived age between healthy controls (HC) and participants with atopic dermatitis (AD), asthma (As), food allergy (FA), allergic rhinitis (AR), or one or more (1+) or two or more (2+) allergic diagnoses at a 3 months and b 1 year. Comparisons were assessed using Wilcoxon rank sum tests for a 3-months: AD, n = 162, p-value = 0.2004; As, n = 84, p-value = 0.4177; FA, n = 57, p-value = 0.2178; AR, n = 82, p-value = 0.9443; 1+, n = 265, p-value = 0.2002; 2+, n = 97, p-value = 0.6144 and b 1-year: AD, n = 179, p-value = 0.0452; As, n = 95, p-value = 0.7817; FA, n = 67, p-value = 0.0105; AR, n = 102, p-value = 0.1738; 1+, n = 309, p-value = 0.0576; 2+, n = 72, p-value = 0.3049. (*) indicate FDR-adjusted p < 0.05. c Top 25 features contributing to fungal-derived age based on model importance. MaAsLin2 analyses of associations with 1-year fungal-derived age are shown for d fungal species (n = 1230), e bacterial species (n = 1230), f bacterial Metacyc pathways (n = 1230), and g batch-corrected metabolites (n = 1054). Models were adjusted for exact age at stool sample collection and sample processing time, with site included as a random effect. Boxplots indicate the median (center line), interquartile range (box bounds), and 1.5 × IQR (whiskers). Points represent individual samples. For Cleveland plots, all features with FDR-adjusted p < 0.25 are shown, with filled in circles indicating FDR-adjusted p < 0.05. Points indicate MaAsLin2 coefficient estimates and error bars represent coefficient estimate ± standard error. Associations were evaluated using MaAsLin2 multivariable linear models with two-sided hypothesis testing. P-values were adjusted for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) correction.

When analyzing genus-level fungal-derived age, Malassezia emerged as the top contributor to fungal-derived age prediction accuracy, followed by unclassified members of Saccharomycetes, the genera Candida, Pichia, Cladosporium, Aspergillus, unclassified member of Agaricales, the genera Rhodotorula, Penicillium, unclassified fungi, unclassified members of Saccharomycetales, the genera Mycena, Hypholoma, Exophiala, Athelia, unclassified member of Ascomycota, and the genera Baespora, Debaromyces, and Naganishia (Fig. 2c and Supplementary Table 1). Contribution scores dropped sharply after the top 10–15 features, suggesting that a core set of fungi define early-life mycobiome maturation. Notably, Malassezia and Saccharomycetaceae, two highly prevalent fungal groups in early life, were among the strongest predictors in this model. Both have been linked to immune modulation and host metabolic processes, aligning with previous findings on ecological assembly and fungal contributions to early-life health^4,22–25. When considering species-level associations with fungal-derived age, fungal species negatively associated with fungal-derived age included a Cladosporium sp., a fungi in the order Agaerocoles, Baeospora myosura, Hypholoma capoides, and a Saccharomycetes sp. (Fig. 2d and Supplementary Table 2).

Next, we analyzed the 1-year bacterial species associated with 1-year fungal-derived age to define other interkingdom factors of the infant gut that may influence the health outcomes of atopic dermatitis and food allergy (Fig. 2e and Supplementary Data File 1). Bacterial species positively associated with fungal-derived age (adj. p < 0.05) were Streptococcus salivarius (SGB8007), Tyzzerella nexilis (SGB4588), Clostridium sp. (SGB6179), and Streptococcus parasanguinis (SGB8071). These bacterial species change in tandem with varying fungal-derived age, suggesting that they contribute to both the ecological environment of the infant gut and the infant’s immune development. In addition, pathways were inferred from metagenomic data by annotating microbial genes according to curated metabolic pathways based on known enzymatic reactions, thereby reflecting the predicted metabolic potential of the community. Using this approach, 58 pathways were significantly associated with fungal-derived age (Fig. 2f and Supplementary Data File 2, adj. p < 0.05). The top bacterial pathways associated with fungal-derived age included stachyose degradation (PWY 6527), D-galactose degradation (PWY 6317), and UDP N-acetylmuramoyl pentapeptide biosynthesis III (PWY 7953).

We then analyzed the 1-year stool metabolites associated with 1-year fungal-derived age (Fig. 2g and Supplementary Data File 3). Metabolites associated with fungal-derived age with an adj. p < 0.05 included methylamine, acetic acid, asparagine, and butyrate. Many of these are known immune-modulating metabolites, with butyrate and acetate additionally demonstrated to alter fungal colonization directly^10,26. Thus, these metabolites demonstrate strong associations with fungal signals within the early-life gut environment.

Distinct infant fungal community types and their link to allergic outcomes

As a complementary step, we determined whether distinct fungal community types, defined by shared taxonomic compositions across CHILD participants, reflected different trajectories of mycobiome maturation and if these trajectories were associated with different health outcomes. Because the gut mycobiota develops over time, we first defined how fungal community types varied between the 3-month and 1-year visits. We applied Dirichlet Multinomial Mixtures (DMM) modeling to identify clusters of similar genus-level profiles within the ITS2 sequencing data²⁷. Using the Laplace approximation to determine model fit, we identified four optimal clusters from 2256 samples, comprising 1121 samples from the 3-month visit and 1293 from the 1-year visit (Supplementary Fig. 2). Cluster definitions were derived from the DMM model, and contributions of taxa to the DMM model were calculated as the mean probability vectors for each component of the identified clusters (Supplementary Table 3). The top contributors included: Cladosporium, Penicillium, and Saccharomycetaceae (Cluster 1); Malassezia, Candida, and Penicillium (Cluster 2); Malassezia, Cladosporium, and Penicillium (Cluster 3), and Saccharomycetaceae (Cluster 4). These clusters may represent fungal states shaped by early life environmental exposures or host-microbe interactions. We therefore characterized their defining features and examined associations with early-life health outcomes.

We first assessed whether the participant’s age at the time of stool sample collection was associated with cluster membership. Two clusters predominated at 3 months (Clusters 1 and 2), and while Clusters 3 and 4 were more common at 1 year (Fig. 3a; Cluster 1 at 3 months: n = 493, Cluster 2 at 3 months: n = 532, Cluster 3 at 3 months: n = 18, Cluster 4 at 3 months: n = 78; Cluster 1 at 1 year: n = 249, Cluster 2 at 1 year: n = 68, Cluster 3 at 1 year: n = 529, Cluster 4 at. 1 year: n = 447), supporting a normative shift in fungal composition with age. To explore the relevance of this variation, we compared participants who did not fall within the standard clusters to those who did. We focused on 1-year samples and compared participants whose fungal profiles more closely resembled 3-month patterns (Clusters 1 and 2, n = 317) to those with typical 1-year profiles (Clusters 3 and 4, n = 976). Demographic characteristics were largely similar between groups (Supplementary Table 4). However, participants with 3-month fungal patterns at their 1-year visit had significantly lower fungal-predicted age (Wilcoxon p < 0.001), even when adjusting for chronological age at sample collection, and decreased fungal Shannon diversity (Fig. 3b, Wilcoxon p < 0.001), indicating delayed or inhibited mycobiome maturation. We therefore refer to these groups as having “immature” (Clusters 1 and 2) and “mature” (Clusters 3 and 4) 1-year mycobiota profiles. Consistent with our earlier association between fungal predictive age and 5-year allergic risk, atopic dermatitis diagnosis at age 5 years was more common among participants with immature fungal profiles (22%) than those with mature profiles (15%) (Fig. 3c, Fisher’s exact test p = 0.007). In comparison, food allergy diagnosis at age 5 years was not significantly associated with fungal profiles (Fisher’s exact test p = 0.645). These findings reinforce that delayed fungal maturation in infancy may be associated with increased risk of atopic dermatitis in particular.

Fig. 3. DMM clusters of the mycobiome across early life and associated bacterial and metabolic features.

a Sankey plot of fungal DMM clusters by visit. At 3 months: Cluster 1, n = 493; Cluster 2, n = 532; Cluster 3, n = 18; Cluster 4, n = 78. At 1-year: Cluster 1, n = 249; Cluster 2, n = 68; Cluster 3, n = 529; Cluster 4 at, n = 447. Cluster colors are white, pink, blue, and yellow for Clusters 1–4, respectively. Orange indicates the immature group and teal indicates the mature group. b Boxplots of fungal-derived age and Shannon diversity comparing immature clusters (Clusters 1 and 2; n = 659) and mature clusters (Clusters 3 and 4; n = 571) at 1 year (Wilcoxon rank sum test p < 0.001). (*) indicate FDR-adjusted p < 0.05. c Mosaic plot of atopic dermatitis (AD) distribution across 1-year clusters. The immature and mature groups were compared using a two-tailed Fisher’s exact test (p = 0.007). Striped bars indicate AD and solid bars indicate no AD. MaAsLin2 analyses comparing immature and mature groups at 1 year are shown for d bacterial species (immature: n = 317; mature: n = 976), e bacterial MetaCyc pathways (immature: n = 317; mature: n = 976), and f batch-corrected metabolites (immature: n = 306; mature: n = 876). Models were adjusted for exact age at stool sample collection and sample processing time, with site included as a random effect. Boxplots indicate the median (center line), interquartile range (box bounds), and 1.5 × IQR (whiskers). Points represent individual samples. For Cleveland plots, all features with FDR-adjusted p < 0.25 are shown, with filled in circles indicating FDR-adjusted p < 0.05. Points indicate MaAsLin2 coefficient estimates and error bars represent coefficient estimate ± standard error. Associations were evaluated using MaAsLin2 multivariable linear models with two-sided hypothesis testing. P-values were adjusted for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) correction.

Bacterial and metabolic signatures of an immature gut mycobiome

We next examined stool bacteria signatures and metabolomic profiles at 1 year across fungal cluster groups. Bacteria associated with differences in the immature gut microbiome with an adj. p < 0.05 included Flavonifractor plautii (SGB15132), Clostridium innocuum (SGB4037), Eggerthella lenta (SGB14809), Clostridium symbiosum (SGB4699), Hungatella hathewayi (SGB4742), Anaerostipes caccae (SGB4529), Erysipelatoclostridium ramosum (SGB6744), Ruthenibacterium lactatiformans (SGB15271), and Eubacterium rectale (SGB4933) (Fig. 3d and Supplementary Data File 4).

Of the pathways associated with the changes in the mature group, 64 were significantly associated with the DMM group (Fig. 3e and Supplementary Data File 5; adj. p < 0.05). Pathways decreased in the mature group included fatty acid beta-oxidation (PWY 66-391) and trans-octadecadienoyl-CoA degradation (PWY 7340), while those increased included D-galactose degradation (PWY 6317). Metabolomic analysis also revealed distinct profiles between cluster groups using linear mixed-effects modeling (Fig. 3f and Supplementary Data File 6). Notably, glycerol was the only metabolite that met an adj. p < 0.05.

The importance of fungi in interkingdom ecological development

Gut fungi coexist alongside dominant bacterial colonizers in the developing infant microbiome. While bacteria are abundant and interact extensively with the host through direct colonization and metabolic activity, fungi represent a much smaller proportion of the total microbial load in the gut²⁸. We therefore evaluated the contribution of fungal taxa relative to bacterial species in predicting infant age.

To explore this, we recalculated a nested cross-validated microbiota-derived age using the combined abundances of bacterial species and fungal species during early life (Fig. 4a and Supplementary Data File 7). The resulting model showed a strong correlation with participant chronological age (Fig. 4a, R = 0.9, p < 0.001), reflecting the predictive power of combined community data. Among the top 20 contributors to the model, 17 were bacterial species, and three were fungi: a Saccharomycetes sp., Malassezia restricta, and Pichia kudriavzevii (Fig. 4b). Notably, the Saccharomycetes sp. and Malassezia restricta ranked within the top 10 most important contributors to the model. These results indicate that, despite their lower abundance, gut fungi contribute to models of microbiota maturation and provide complementary information to bacterial features, potentially reflecting distinct structural or ecological roles within the gut.

Fig. 4. Characterizing interkingdom dynamics.

a Random Forest bacteria- and fungi-derived microbial age of stool samples plotted against chronological age, with linear regression lines of best fit and 95% confidence intervals (Spearman R = 0.9, p < 0.001). Density is represented by (+) red and (−) blue. b Top bacterial species and fungal genera contributing to fungal-derived age based on model importance. Orange indicates fungi and gray indicates bacteria.

Discussion

Infant mycobiome studies have expanded rapidly, offering insights into early-life fungal colonization^3,4,29–31. Yet many are constrained by small cohorts or a lack of multi-omic resolution. In contrast, our study integrated ITS2 amplicon sequencing, shotgun metagenomics, and metabolomics in a large, longitudinal birth cohort to characterize the early-life gut mycobiome alongside bacterial and metabolic profiles. While early life is widely recognized as a critical window for bacterial colonization and immune development, fungal dynamics during this period are only beginning to be elucidated and may offer novel opportunities for intervention^1–3,32.

We observed clear temporal patterns in fungal composition. Saccharomycetaceae and Malassezia were dominant during early life, with Saccharomycetaceae increasing and Malassezia decreasing over the first year. These shifts may reflect changes in environmental exposures, such as breastfeeding, a known determinant of early gut ecology and fungal colonization^33,34. Notably, Malassezia, Saccharomycetaceae, Pichia, and Candida emerged as potentially important contributors to microbiota maturation, with Malassezia consistently ranked among the strongest temporal markers, even relative to bacterial taxa. Interestingly, a recent study characterizing mycobiome maturation and clinical determinants of mycobiome compositional variation during the first 2 years of life showed that fungal richness increased and mycobiome composition changed in a similar ordered pattern during the first 2 years of life³⁵. Predominant taxa included Candida albicans, Saccharomyces, and Malassezia restricta. Key influences reported included antibiotic exposure and breastfeeding status contributed to time-specific mycobiome compositional variation. Our studies report similar findings, and the aforementioned study provides additional insight into the early life contributors to mycobiome development.

Early life represents a critical window for gut mycobiome assembly, during which fungal communities are low in diversity but disproportionately influential for immune development^4,22. Infant gut mycobiomes are consistently dominated by a limited set of taxa, including Candida, Saccharomyces, Malassezia, Rhodotorula, Cladosporium, and Penicillium, with composition strongly shaped by delivery mode, antibiotic exposure, diet, and environmental contact^3,4. Among these, Candida spp., particularly C. albicans, often dominate early infancy and expand following antibiotic-mediated bacterial depletion, with multiple longitudinal cohorts linking increased early-life Candida abundance to subsequent development of asthma, eczema, and food allergy³⁰. In contrast, Saccharomyces spp. are typically transient and diet-associated, and their presence has been correlated with greater overall microbial diversity and reduced allergic risk, potentially reflecting a more balanced microbial ecosystem rather than a direct protective effect⁴. Environmental fungi such as Cladosporium are frequently detected during early infancy but decline with age and enrichment of these taxa has been associated with later allergic sensitization, likely reflecting increased early environmental fungal exposure during a period of immune immaturity³. Malassezia, traditionally considered skin-restricted, is increasingly detected in the infant gut and is enriched in infants who later develop atopic dermatitis, supporting a model of coordinated fungal-driven immune priming across epithelial barriers³⁶.

We previously reported that reduced bacterial maturation at 1 year is linked to increased risk of multiple allergic outcomes, including atopic dermatitis, food allergy, allergic rhinitis, and asthma by age 5 years⁹. In the current study, reduced fungal maturation at 1 year was specifically associated with atopic dermatitis and food allergies only. This narrower association may reflect fundamental differences in host–fungal interactions. As eukaryotes, fungi engage distinct immune recognition pathways, including Dectin-1 and CARD9 signaling, which are implicated in epithelial barrier integrity and allergic sensitization^37,38. Given that both atopic dermatitis and food allergy often originate from epithelial barrier dysfunction, our findings suggest that fungal maturation may influence immune development through the gut–skin axis. Notably, atopic dermatitis frequently precedes food allergy, and early-life compromise of the skin barrier is a known risk factor for food sensitization^39,40. These results support a potential role for gut fungal maturation in modulating epithelial and immune responses relevant to allergic disease.

Despite lower abundance, fungal cells may play outsized ecological and immunological roles due to their larger size, filamentation potential, and ability to interact directly with host tissues^41,42. For example, the genus Malassezia was a primary biomarker of the immature clusters characterized within this study. Malassezia is more commonly studied on the skin; however, its consistent presence and predictive importance in our infant stool samples highlight its underappreciated role in the gut. Malassezia is evolutionarily adapted for life on host surfaces, with genomic traits such as lipase, protease, and phospholipase production that may facilitate survival at epithelial interfaces^43,44. Malassezia’s early-life abundance, followed by a sharp decline, may reflect a transitional ecological role during gut colonization. Together with this study, the concurrent manuscript by Bernardes et al. provides complementary evidence that early-life fungal colonization is an important and previously underappreciated feature of immune development and atopic disease¹¹. While our work identifies persistent gut colonization of Malassezia spp. as a hallmark of delayed fungal microbiome maturation associated with increased risk of atopic dermatitis, Bernardes et al. extend these findings by demonstrating that antibiotic exposure in infancy promotes expansion of Malassezia. They further demonstrate, using gnotobiotic mouse models, that Malassezia can contribute to immune dysregulation and enhanced airway inflammation. Together, these studies connect early-life disruption of the infant mycobiome with altered immune development and allergic disease susceptibility, providing ecological, clinical, and mechanistic support for the infant mycobiome as a promising target for strategies aimed at preventing pediatric allergic diseases.

We observed metabolomic differences associated with fungal maturation, including changes in butyrate, acetic acid, and pyroglutamic acid, all metabolites with well-established immunomodulatory functions^45,46. The overlap between fungi and bacterial-associated metabolites in our analysis suggests interkingdom coordination in shaping the gut’s metabolic and immunological environment. This is consistent with emerging evidence that fungal–bacterial co-colonization and competition help structure microbial communities and influence host responses⁴⁷. Notably, we also observed trends of an increase in taxa and Metacyc pathways involved in galactose metabolism (Eubacterium rectale, D-galactose degradation (PWY 6317)), a pathway previously linked to allergic disease alleviation⁴⁸. Dietary galactose and its derivatives can be metabolized by gut microbes, altering the composition and activity of the microbiota and potentially acting as a prebiotic that shapes microbial communities and their metabolic outputs.

Broad-spectrum interventions such as fecal microbiota transplantation may ultimately support both fungal and bacterial restoration, but safety, ethical, and logistical challenges currently limit their use in infants. Although most microbiota-targeted therapies to date have focused on bacteria, emerging work on fungal-based interventions highlights a promising yet underexplored area⁴⁹. Developing safe and effective strategies to support optimal early-life mycobiome development remains an important frontier for future research.

Our study has several limitations common to longitudinal cohort designs with integrated microbiome analyses. While powerful for uncovering population-level insights, such studies often face confounding variables and biased sampling due to the tendency for more financially and temporally resourced individuals to participate. Future cohort studies would also benefit from a more granular characterization of atopic dermatitis phenotypes by location and clinical subtype. Evidence shows that fungal sensitization, particularly to Malassezia spp., is especially common in head-and-neck dermatitis (HND)^50–52. Intriguingly, HND also differs in its clinical presentation depending on age of onset, wherein adolescent-onset HND typically involves only head-and-neck areas and adult-onset disease frequently co-occurs with widespread atopic dermatitis. Given these distinctions, future research would do well to stratify participants by both anatomic distribution and AD subtype, enabling more precise mapping of sensitization profiles and their clinical implications. Additionally, although not investigated within this study, the role of postnatal timing for the introduction of solid foods is a critical future area for investigation and ongoing analysis regarding mycobiome development during early life. Finally, the analysis of bacterial-fungal associations via the use of co-occurrence network analysis represents a promising future direction with larger datasets through which the current results can be extended.

As with all sequencing-based microbiome studies, technical and biological constraints limit detection of low-abundance taxa, particularly for fungal communities, which are challenged by low biomass and incomplete reference databases. Finally, limited stool sample volumes restricted analyses to sequence-based approaches, precluding culture-based validation. Future studies incorporating expanded reference databases, improved sequencing technologies, quantitative approaches to estimate absolute abundance, and integrated culture-based and multi-omic methods will be important for refining ecological and causal inference.

In sum, our findings reinforce the view that fungi are active, informative members of the early-life gut ecosystem. Their consistent association with microbiota maturation and allergic outcomes suggests that fungi influence health trajectories in ways not captured by bacterial data alone. By characterizing microbial communities across fungal, bacterial, and metabolic domains, we aim to support more holistic strategies for promoting healthy immune development and reducing allergy risk in childhood.

Methods

CHILD Study participants

Ethical approval for the CHILD Cohort Study, including the oversight of the CHILD biological samples and the CHILD database (CHILDdb), was obtained from the local Research Ethics Board of each study site: the University of British Columbia, the University of Alberta, the University of Manitoba, the Hospital for Sick Children, and McMaster University. The Research Ethics Board Number is H07-03120. Informed written consent was obtained from the parents or legal guardians of all participating children.

This project significantly expands the understanding of the early life gut mycobiome within participants within the CHILD Study by providing a novel and independent measure of microbiota changes by analyzing the mycobial component. This multi-center longitudinal, prospective, general population birth cohort study follows infants from pregnancy to age 5 years and beyond. Briefly, enrolment for CHILD began in 2008. Recruitment closed in 2012, including a total of 3621 pregnant people from four cities (Vancouver, Edmonton, Winnipeg, Toronto) across Canada enrolled along with eligible infants (n = 3455) that had no congenital abnormalities and were born at a minimum of 34 weeks of gestation⁵³. Informed consent was obtained from parents/legal guardians during this study, and detailed information was collected using a combination of questionnaires and in-person clinical assessments. This study included stool samples with corresponding shotgun metagenomics data and ITS2 amplicon sequencing data (n = 2256, Supplementary Fig. 3). The composition of the biological sex of subjects included n = 526 3-month samples from male participants, n = 594 3-month samples from female participants, n = 625 1-year samples from male participants, and n = 667 1-year samples from female participants (based on family/participant reporting). Biological sex and gender were not adjusted for in the study design.

We derived variables indicating whether participants were diagnosed with a condition of interest or had no conditions up and through their 5-year evaluation and used multivariable conditional logistic regression (stratified by study center) to evaluate the influence of early-life and familial exposures, including biological sex, mode of delivery at birth, breastfeeding status at the age of 6 months, upon atopic condition development by the age of 5 years. Missing data were considered missing completely at random and individuals were removed from the multivariable analysis if they had a missing value in any covariates.

The primary clinical outcome included within this study was atopic dermatitis at age 5 years (clinically binned as Yes/Possible/No). This was done using history and physical examination combined with SPT by an expert study physician at the clinical assessment at 5 years of age based on a published approach⁵³. For this study, children were considered to have allergic diseases only if the expert physician’s recorded diagnosis was a definitive “Yes” to whether a participant had atopic dermatitis, asthma, food allergy, and/or allergic rhinitis. Non-allergic controls were limited to children with “No” responses for 5-year diagnoses, negative allergen SPTs at 1, 3, and 5 years, and no history of wheezing at 1, 3, and 5 years for the comparison of derived microbial age of allergic diagnoses to the healthy control subset as we previously described⁹.

All infants enrolled in the CHILD protocol were administered an SPT at their 1-, 3-, and 5-year scheduled visits. Children were then diagnosed with IgE-mediated allergic sensitization (also referred to as atopy) based on SPT to multiple common food and environmental inhalant allergens, using ≥2 mm average wheal size as indicating a positive test relative to the negative control. The allergens tested at all 1-, 3-, and 5-year visits include cat hair, the German cockroach, Alternaria tenuis, house dust mites (Dermatophagoides psteronyssinus and Dermatophagoides farnae), dog epithelium, cow’s milk, peanut, egg white, and soybean. In addition to these, participants at 3- and 5-year visits were tested with Cladosporium, Penicillium, Aspergillus fumigatus, trees, grass, weeds, and ragweed. Glycerin and histamine served as the negative and positive controls, respectively. When analyzing only atopic dermatitis for the cluster analysis, children were considered allergic only if the expert physician’s recorded diagnosis was a definitive “Yes” to whether a participant had atopic dermatitis and participants were considered controls only if the expert physician’s recorded diagnosis was a definitive “No” to whether a participant had atopic dermatitis.

Stool sample collection

Sample collection and sequencing were performed as previously described⁵⁴. Briefly, stool samples from diapers were collected at a home visit at the mean (SD) of 3.8 (1.1) months and at a clinic visit at 12.5 (1.6) months. Samples were briefly refrigerated (4 °C) and then aliquoted into four cryovials (2 mL each) using a stainless steel depyrogenated spatula. They were then frozen at −80 °C, with time between collection and long-term storage recorded for statistical adjustment.

Extraction, library preparation, and ITS2 amplicon sequencing

Sample DNA extraction, library preparation, and ITS2-based amplicon sequencing of fungal rDNA from stool samples were performed by Diversigen (Minneapolis, MN, USA). Specifically, samples were extracted with MO Bio PowerFecal (Qiagen), automated for high throughput on QiaCube (Qiagen), with bead beating in 0.1 mm glass bead plates. Samples were quantified via qPCR using primers for ITS2 (5.8SR, ITS4) between the 5.8S and 26S rRNA subunits⁵⁵. Samples were prepared with a protocol derived from ref. ⁵⁶, using KAPA HiFi Polymerase to amplify the ITS2 region (5.8SR, ITS4)⁵⁷. Libraries were sequenced on an Illumina MiSeq using paired-end 2 × 250 reads and the MiSeq Reagent Kit v3. DNA sequences were filtered for low quality (Q-Score < 30) and length (<50 bases), and adapter sequences were trimmed using “cutadapt”⁵⁸. Fastq files were converted to a single fasta of stitched reads using “shi7”⁵⁸. Raw reads were processed in QIIME 2 2022.2⁵⁹. Taxonomic assignment was performed using the UNITE fungal ITS database (version 10.0) within the QIIME2 framework.

Additionally, the “phyloseq” package was used to pre-process the ITS2 taxonomy table⁶⁰. Unknown amplicon sequence variants were removed, samples were pruned to 4 k sequencing reads, and taxa were aggregated to the species and genus-level, depending on the analysis. Some ITS2-derived annotations were resolved only to higher taxonomic ranks due to limitations in fungal reference databases; these annotations were retained as reported by the classifier and standardized throughout the manuscript for consistency. Statistical tests were applied to the log-transformed normalized relative abundance of fungi to identify fungal taxa significantly associated with outcomes of interest. Models were only applied to fungal genera detected in at least 10% of used samples (42 genera), the default setting in the MaAsLin2 package. MaAsLin2 adds a pseudocount of half the minimum genera levels detected before the log transformation relative abundances. P-values were corrected using the Benjamini–Hochberg approach; results with adj. p < 0.05 were considered significant and presented as so.

We processed our ITS2 amplicon data using QIIME 2, first running qiime quality-control decontam-identify to flag potential contaminant ASVs, and then removing them with qiime feature-table filter-features and filter-seqs based on decontam scores⁶¹. After this decontamination step, we rarefied each sample to 2000 reads using QIIME 2’s rarefaction tools to normalize sequencing depth. For diversity analyses (both α- and β-diversity), we used the ASV-level feature table to preserve maximum resolution. For compositional and taxonomic analyses, we aggregated ASVs up to the genus or species level based on their QIIME 2-assigned taxonomy, allowing clearer biological interpretation.

Shotgun metagenomic sequencing

As described previously⁵⁴, shotgun metagenomic sequencing data were generated by Diversigen (Minneapolis, MN, USA). Again, the bioBakery 3 pipeline was used to map sequences and classify sequences into each sample’s taxonomy (species level) and functional features. However, in this case, MetaPhlAn 4 was used for taxonomic classification and HUMAnN 4 for functional profiling. The “phyloseq” package was again used to preprocess the metagenomic taxonomy table⁶⁰.

Models were only applied to species and MetaCyc pathways detected in at least 10% of used samples were tested (72 species and 347 pathways, respectively), the default setting in the MaAsLin2 package. MaAsLin2 adds a pseudocount of half the minimum species or MetaCyc pathway level detected before the log transformation relative abundances. P values were corrected using the Benjamini–Hochberg approach and results with an adj. pval<0.05 were considered significant and presented as so.

Longitudinal trends in the mycobiome

Line graphs depicting the longitudinal changes in fungal taxa were generated using “ggplot” with the geom_smooth() function. All samples from participants were aggregated, with age at time of collection used as the time of mycobiome profiling. Genera were transformed using the “microbiome” package and transform() function for compositional calculation. Genera and species with >2% relative abundance and >10% prevalence were included.

Maturation metrics

The fungal and interkingdom-predicted ages were estimated using a nested fivefold cross-validated random forest regressor, with the same approach reported in our previous paper⁹. However, for this study, the interkingdom-predicted age was derived from species-level bacterial and genus-level fungal relative abundance data, genus-level fungal data by itself, and species-level bacterial and fungal relative abundance data. We used the “randomForest,” “mlbench” (Leisch & Dimitriadou, 2024), and “caret” packages in R to generate the predicted age (gut maturation) using the nested fivefold cross-validated random forest model with an mtree value of 500^62,63. Within each fold of the cross-validated analysis, the hyper-parameters of the random forest model were tuned given a grid search space using nested fivefold cross-validation. The microbial age was the combination of the predicted value of each holdout set from each cross-validated random forest regression. The importance of features was the average of importance scores from all models in the nested cross-validated regressor.

ITS2 DMM modeling

To carry out DMM modeling, we utilized the “DirichletMultinomial” package in R⁶⁴ with a seed of 1 and a limit of 10 clusters. DMM modeling accounts for several differences between samples: size, sparsity, diversity, and skewed nature; each community is described as a vector of taxa probabilities using the count data from the samples⁶⁴. The vectors are generated from a finite number of Dirichlet mixture components with different hyperparameters, which cluster communities to form groups of communities with a similar composition, known as “enterotypes”⁶⁴. Following this, we determined the optimal number of clusters for the fungal groups by identifying the number of clusters with the lowest model fit, accounting for complexity using Laplace approximation²⁷. Contributions of taxa to the DMM model were calculated as the mean probability vectors for each component of the identified clusters.

Metabolomics pre-processing and analysis

The metabolomics pre-processing was performed as previously⁹. Briefly, metabolic profiles were created from the same sequenced stool samples as the shotgun metagenomic and ITS2 amplicon sequencing at The Metabolomics Innovation Centre (TMIC) in Edmonton, Alberta, using two separate assays. Targeted nuclear magnetic resonance (NMR) analysis of 31 metabolites was performed across 62 batches. Targeted liquid chromatography with tandem mass spectrometry (LC-MS/MS) analysis of 590 metabolites was performed using TMIC’s Microbiome Metabolism (MEGA) assay across 27 batches⁹. NMR and LC-MS/MS precision were confirmed to be <5 and <10% coefficients of variability (CV), respectively. Additionally, overlapping metabolites detected by both methods were cross-checked to confirm the accuracy of the reported concentration values.

Detailed NMR and LC-MS/MS analysis methods are in our previous publication⁹. All 31 metabolites from the NMR analysis were kept for downstream analysis. Of the 590 metabolites targeted in the LC-MS/MS analysis, we excluded metabolites detected below the detection limit in more than 80% of samples (meaning they were present in less than 20% of the samples). This resulted in the exclusion of 244 metabolites. Remaining metabolite concentrations below the limit of detection were imputed with a value of one-half the minimum concentration for each metabolite and log-transformed. An additional 132 low-variance metabolites based on standard deviation (log(SD) less than −5) were excluded. Technical sample outliers were detected via PCA analysis, followed by the quantification of local outlier factor (lof), using the “stats” and “dbscan” packages, respectively. Samples with a lof greater than 5 were excluded (3 LCMS samples and 2 NMR samples). Before any downstream analysis, the resulting NMR and LC-MS/MS were batch-corrected using the “ComBat” package. This reduced the effect of the batch from R² = 0.11 to R² = 0.017 in the LC-MS/MS dataset and reduced the effect of the batch from R² = 0.13 to R² = 0.004) in the NMR dataset. Batch-corrected datasets containing a total of 245 metabolites were merged for downstream analyses.

Ethical oversight

This study was overseen and approved by the Research Ethics Boards (CREBs) of University of British Columbia, University of Manitoba, University of Toronto, McMaster University, BC Children’s Hospital, The Hospital for Sick Children, and Simon Fraser University under the REB Number H07-03120. These reviews and approval were in accordance with the requirements of the Tri-Council Policy Statement: Ethical Conduct for Research involving Humans (TCPS2, 2018).

Statistics and reproducibility

The “Maaslin2” package was used to perform linear mixed-effects models (MaAsLin2 function)⁶⁵ with study center location as a random effect and adjusting for age at stool sample collection and processing time to examine the association between microbial community structure at 1 year of age. P-values were corrected using the Benjamini–Hochberg approach, and results with adj. p < 0.05 were considered significant and presented as such. Additionally, we highlighted results which passed a threshold of adj. p < 0.05. For analyses of differences between groups for chronological age and predicted age metrics, Wilcoxon tests were used to test the differences between DMM clusters and cluster groups.

No statistical method was used to predetermine sample size. Data were only excluded from the analysis if they did not meet the criteria outlined in the Child Study participants and metagenomic sections above. The experiments were not randomized. The Investigators were not blinded to allocation during experiments and outcome assessment. Blinding was not feasible within this observational longitudinal cohort study because investigators and clinicians were responsible for collecting clinical metadata, administering questionnaires, and performing participant assessments during scheduled study visits. Participant group assignment was determined retrospectively based on physician-diagnosed allergic outcomes and clinical assessments rather than experimental intervention allocation. Laboratory-based sequencing and metabolomic analyses were performed using standardized protocols independent of participant outcome status, minimizing the potential for analytical bias.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization, C.H., D.L.Y.D., C.P., and S.E.T.; Methodology, C.H., D.L.Y.D., and C.P.; Investigation & formal epidemiological, metagenomic, and metabolomic analysis, C.H., D.L.Y.D., and C.P.; Visualization, C.H. Data oversight—statistical analyses, C.H., D.L.Y.D., and C.P.; Unrestricted access to all data, C.H., D.L.Y.D., C.P., and S.E.T.; First draft, C.H., D.L.Y.D., C.P., and S.E.T.; Review & editing, C.H., D.L.Y.D., T.J.M., P.J.M., E.S., A.L.K., P.S., M.B.A., C.P., and S.E.T. Funding acquisition, S.E.T.; Resources, M.B.A., P.S., and S.E.T. All authors agreed to submit the manuscript, read and approved the final draft and take full responsibility of its content, including the accuracy of the data and its statistical analysis.

Peer review

Peer review information

Nature Communications thanks Tobias Hohl and the other, anonymous, reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

C.H. held funding by the University of British Columbia John Richard Turner Fellowship in Microbiology, the President’s Academic Excellence Initiative PhD Award, and the University of British Columbia Four Year Doctoral Fellowship (4YF) during the majority of this study. D.L.Y.D. also held funding by a Canadian Institute of Health Research Frederick Banting and Charles Best Canada Graduate Scholarship Doctoral Award (CIHR CGS-D) and the University of British Columbia 4YF during this study. M.B.A. holds a Tier 2 Canada Research Chair in the Developmental Origins of Chronic Disease and is a Fellow if the CIFAR Humans and the Microbiome Program. Gut microbiota analysis at the University of British Columbia was funded by CIHR (OGB-185749), AllerGen NCE and Genome Canada and Genome British Columbia ([274CHI] and [EC1-144621] S.E.T.). P.S. holds a Tier 1 Canada Research Chair in Pediatric Asthma and Lung Health. S.E.T. holds a Tier 1 Canada Research Chair in Pediatric Precision Health, the Aubrey J. Tingle Professorship of Pediatric Immunology, and the O’Sullivan Family Hospital Chair in Precision Health Research at BC Children’s Hospital. We further acknowledge the support of BC Children’s Hospital Research Institute and Foundation and the Provincial Health Services Authority.

Data availability

The shotgun metagenomic data generated in this study have been deposited in the NCBI database under accession code PRJNA838575. The fungal sequencing data generated in this study have been deposited under accession code PRJNA1368998. The metabolomics data generated in this study are available in the MetaboLights database under accession code MTBLS7919. The informed consent obtained from the CHILD participants, in addition to the CHILD Inter-Institutional Agreement, which has been executed between the five Canadian institutions responsible for the study, govern the sharing of CHILD data. To facilitate recruitment and retention, participants were reimbursed for their time in the form of gift cards and similar reimbursements. Data described in the manuscript are available by registration to the CHILD database (https://childstudy.ca/childdb/) and the submission of a formal request. All reasonable requests will be accommodated. More information about data access for the CHILD Cohort Study can be found at https://childstudy.ca/for-researchers/data-access/. Researchers interested in collaborating on a project and accessing CHILD Cohort Study data should contact child@mcmaster.ca.

Code availability

Code for study can be found in the Turveylab GitHub (https://github.com/turveylab/).

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-74418-w.