Environmental and skin–nasal microbiome variation in South African children with atopic dermatitis Authors: Yang Chen, Anna Dan Nguyen, Nonhlanhla Lunjani, Gillian Ndhlovu, Drishti Kaul, Britta De Pessemier, Teruaki Nakatsuji, Murunwa Madzinga, Evgeniya Sovershaeva, Carol Hlela, Michael Levin, Avumile Mankahla, George Hightower, Chris Callewaert, Rob Knight, Richard L Gallo, Christopher L Dupont, Felix Dube Published: 2025 ABSTRACT AbstractBackgroundAtopic dermatitis (AD) in early childhood is associated with microbial dysbiosis. Skin and nasal microbiomes have been linked to AD severity; this relationship has not yet been studied in an African cohort. Here, we aimed to explore how urban and rural stratification, disease severity, and inter-site bacterial overlap shape the skin and nasal microbiomes of South African children with AD.MethodsChildren were recruited from urban Cape Town (CT) and rural Umtata (UM), South Africa. We profiled the skin and nasal microbiomes of 183 children (84 healthy controls and 99 with AD; ages 9–37 months), totaling 462 samples, including both lesional and non-lesional skin sites in children with AD, in a cross-sectional study design. Using 16S rRNA V4–V5 sequencing for its accessibility, we applied random forest (RF) models to classify AD status based on amplicon sequence variants (ASVs) and analyzed microbiome composition and diversity by region.ResultsWe found that RF models could predict AD status using both skin and nasal microbiomes (AUCs: skin = 0.69–0.79; nasal = 0.65), strongly driven by both Streptococcus and Staphylococcus. The correlations between skin and nasal microbiomes were significantly stronger in children with AD compared to controls, with higher correlations observed in rural UM (healthy r = 0.45 to AD r = 0.67) compared to urban CT (healthy r = 0.27 to AD r = 0.65). The skin microbiome diversity was higher in children from rural UM with healthy skin than in those from urban CT (p = 0.004). However, children with AD in both groups showed significant alterations in their microbiome, with those in rural UM exhibiting greater beta diversity changes (p = 0.001–0.002) than their urban CT counterparts (p = 0.002–0.349).ConclusionIn South African children with AD, skin–nasal microbiomes reflect shared reservoirs, and differences in the AD microbiome were observed between environmental regions. These findings highlight the need for geographically diverse studies incorporating skin and mucocutaneous sampling to better understand pediatric AD.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12866-025-04589-x.Keywords: Atopic Dermatitis, Eczema, Cchildhood, Microbiome, Skin, Nasal, Staphylococcus, Streptococcus Introduction Atopic dermatitis (AD) is a chronic inflammatory skin disease that typically begins in early childhood and affects up to 20% of children worldwide [1], with rising rates in many African countries [2–4]. The International Study of Asthma and Allergies in Childhood estimated a 23.3% lifetime prevalence of AD among African children aged 6–7, significantly higher than the global average of 14.2% [3]. It is well established that skin structure, barrier function, microbial composition, and immune responses vary across populations, with differences reported among African, Asian, and Western cohorts in both healthy skin and in AD [5–8]. Therefore, region-specific microbiome studies are crucial for elucidating how geographic and environmental contexts influence the pathophysiology of atopic dermatitis. The skin microbiome plays a central role in AD by modulating immune responses and maintaining barrier integrity [9–13]. Colonization by Staphylococcus is known to promote inflammation through the production of cytotoxins and proteases, which impair the skin barrier and activate type 2 immunity [11, 14]. Pediatric AD is distinct in skin physiology and microbial colonization from adult AD, reflecting changes in skin with age [15]. In children, AD is often characterized by increased Streptococcus abundance [15] and variable S. aureus colonization, whereas adult AD is more consistently driven by S. aureus or other Staphylococcus species such as S. epidermidis [11, 16]. Beyond the skin, evidence suggests that the nasal microbiome may also play a role in AD. The nasal cavity contains colonizing microbes commonly found on the skin and may serve as a reservoir for AD-associated taxa [17, 18]. Individuals with AD are over five times as likely to be nasally colonized by S. aureus as healthy controls [19]. Longitudinal data from a large birth cohort have further linked early and frequent nasal colonization with S. aureus during infancy with increased risk and severity of AD [20]. It has also been observed that changes in the skin and nasal microbiome of Staphylococcus species occurred following treatment of AD with Dupilumab [21, 22]. Totté et al. previously identified that both the skin and nasal microbiomes are associated with AD severity in children, particularly driven by Staphylococcus [23]. However, the study did not include healthy controls or non-lesional skin samples, and the association was observed in a Dutch cohort. Here, we expand upon this knowledge through our unique dataset of skin and nasal microbiomes of 183 South African toddlers (84 healthy, 99 with AD) using 16S rRNA V4–V5 sequencing and analysis of 462 samples (Fig. 1). The complete demographic and cohort sample numbers of children in this study are described as within the summary table (Table 1). We employed 16S rRNA gene sequencing for its accessibility and scalability. Recognizing the limited resolution of 16S at the species and strain levels, we applied random forest models to classify AD status from amplicon sequence variants (ASVs). Leveraging a geographically stratified large cohort, including both lesional and non-lesional sites from urban and rural regions, we explored how geography, disease severity, and skin/nasal microbial overlap are associated with the microbiome in early-childhood AD. Graphical abstract. Environmental context influences AD severity and microbiome composition across skin and nasal sites in South African Children. Skin and nasal swabs were collected from 183 South African children aged 9–37 months, including 84 healthy skin controls and 99 with AD. Lesional and non-lesional skin sites were sampled from AD participants, with non-lesional swabs taken from the mid-back. Healthy controls were swabbed at the same mid-back location. Nasal swabs were collected from all participants. AD severity was assessed using the oSCORAD scale. Samples were collected from urban CT (n = 199) and rural Umtata (n = 263), totaling 462 samples (282 skin and 180 nasal) for 16S rRNA (V4–V5) sequencing. The diagram summarizes the study design and hypothesized microbial connectivity between skin and nasal sites in AD Table 1Demographic and sample numbers of study participants by cohort and geographic location. Summary of participant demographics and biological samples collected from children with atopic dermatitis (AD) and healthy controls across two study sites in South Africa. Data are presented as mean ± standard deviation for age, and as counts for sex distribution (M/F = male/female) and sample types. Lesional skin samples were collected only from AD participants. Nonlesional skin and nasal (anterior nares) samples were collected from both AD and control participants. Cape Town (CT) represents the urban cohort, while Umtata (UM) represents the rural cohort. The study included 183 unique children (99 with AD, 84 controls) with a total of 462 biological samples collected across three anatomical sites. NR refers to sex not recorded. Boldface rows represent subtotals by skin disease status and from the total cohort Demographic and sample numbers of study participants by cohort and geographic location. Summary of participant demographics and biological samples collected from children with atopic dermatitis (AD) and healthy controls across two study sites in South Africa. Data are presented as mean ± standard deviation for age, and as counts for sex distribution (M/F = male/female) and sample types. Lesional skin samples were collected only from AD participants. Nonlesional skin and nasal (anterior nares) samples were collected from both AD and control participants. Cape Town (CT) represents the urban cohort, while Umtata (UM) represents the rural cohort. The study included 183 unique children (99 with AD, 84 controls) with a total of 462 biological samples collected across three anatomical sites. NR refers to sex not recorded. Boldface rows represent subtotals by skin disease status and from the total cohort Results Random forest models classify AD status from skin and nasal microbiomes Using 16S relative abundances of ASV bacterial features from both skin and nasal samples, we used RF models to evaluate whether bacterial composition can distinguish AD status [24, 25]. As expected, skin and nasal microbiomes were very highly distinguishable, with high classification accuracy (AUC = 0.95 ± 0.02) (Fig. 2A, left). Dolosigranulum was the top-ranked feature distinguishing skin from nares, with an abundance count log-fold-change (LFC) associated with nares (+), consistent with its well-documented presence in the nasal cavity [26, 27] (Fig. 2B). Random forest classifiers distinguish AD status using skin and nasal microbiome. A ROC curves from RF models trained on 16S ASV-level relative abundances. Left: Classifier distinguishes skin from nares samples (AUC = 0.95 ± 0.03). Middle: Classifiers differentiate ADL, ADNL, and H samples (AUCs = 0.82 ± 0.01, 0.76 ± 0.01, and 0.70 ± 0.03, respectively). Right: Nares microbiome predicts AD status (AUC = 0.68 ± 0.04). Shaded regions denote ± 1 standard deviation from 5-fold cross-validation. B Top ASV-level features contributing to classification across all models. Heatmap shows the rank of each ASV within the top 10 most important features per model (lower ranks, 1 = highest, and higher color intensity indicate greater importance. ASVs belonging to Staphylococcus and Streptococcus were consistently among the most predictive taxa within AD classifications. (+) and (-) values in the heatmap represent log fold changes; (+) is associated with the first group, (-) with the second group RF models could accurately predict AD status, particularly when comparing lesional AD (ADL) and healthy (H) skin (AUC = 0.79 ± 0.02), and AD non-lesional (ADNL) vs. ADL (AUC = 0.75 ± 0.01), as well as ADNL vs. H (AUC = 0.69 ± 0.08) (Fig. 2A, middle). Across all skin disease classification tasks, Streptococcus ASV-1 was the top-ranked feature (lower rank values and higher color intensity reflect greater feature importance) for H vs. ADL and ADNL vs. ADL comparisons, while Staphylococcus ASV-1 ranked top for ADNL vs. H (Fig. 2B). For both ASVs, the LFC values were differential towards the phenotypically more severe groups (ADL > ADNL > H). A key finding in this study was that models trained on nasal microbiomes only were also able to predict AD status with moderate accuracy (AUC = 0.685 ± 0.06) (Fig. 2A, right), aligning with previous findings that disease-associated bacterial signatures are not limited to the skin but are also detectable in the nasal cavity [21, 23]. Notably, Streptococcus ASV-1 emerged as the top predictive feature with LFC towards H (-), followed by Staphylococcus ASV-1 with LFC towards AD (+) (Fig. 2B). Additionally, Dolosigranulum ASV-1 and Dolosigranulum ASV-2 exhibited very low rank feature importances in healthy versus AD comparisons, consistent with the known antagonistic activity of Dolosigranulum against S. aureus [28, 29]. To further distinguish these findings across geographic regions, we performed RF classifications using skin and nasal microbiomes from the UM and CT cohorts separately (Fig. S1). Skin-based models from UM exhibited slightly stronger stratification of AD status (AUCs = 0.75–0.87), particularly for ADL vs. H and ADNL vs. H comparisons, compared to CT (AUCs = 0.70–0.85), suggesting a more detectable AD microbiome disruption in UM that enhances classification accuracy. Nares-based models showed strong predictive ability in the CT cohort (AUC = 0.83 ± 0.07) and moderate predictive ability in the UM cohort (AUC = 0.61 ± 0.10). In addition to Random Forest, we also tested other ensemble-based machine learning methods, including XGBoost, LightGBM, and support vector machine (SVM), which all showed similar AUC trends across classification tasks (Fig. S2). AD is associated with increased skin–nasal bacterial overlap and correlation We examined bacterial overlap between the skin and nares by analyzing ASV sharing within individuals with paired samples. Children with AD demonstrated noticeably higher skin–nasal ASV overlap compared to healthy controls (Fisher’s exact test; p = 8.1e-07, OR = 1.71). Among children with AD, a total of 372 ASVs were shared between skin and nasal samples, whereas 277 ASVs were unique to the skin and 61 were unique to the nares (Fig. 3A, left). In contrast, healthy individuals exhibited fewer shared ASVs (n = 272), with a greater number of taxa uniquely detected in either the skin (n = 347) or nares (n = 75) (Fig. 3A, right). AD is associated with increased skin–nasal bacterial overlap and correlation. A Venn diagrams showing the number of unique ASVs detected exclusively in the skin (purple), exclusively in the nares (orange), or both sites (pink), aggregated across all skin and nares samples from AD (left) and H (right) individuals. B Scatter plots comparing log-transformed mean relative abundance correlations of ASVs shared between skin and nares in CT healthy, CT AD, Umtata healthy, and Umtata AD. Each point represents an ASV with shared prevalence in both body sites across disease status and region-stratified groups. Pearson correlation coefficients (r) and corresponding p-values reflect the concordance in relative abundance between skin and nasal sites To evaluate the relative abundance correlation of shared ASVs, we compared log-transformed mean relative abundances of shared ASVs between skin and nares across four groups: CT healthy, CT AD, UM healthy, and UM AD. We observed significant positive correlations between skin and nasal ASV abundances in all groups. However, the strength of these correlations was noticeably higher in children with AD than in healthy children from both regions, and in the UM rural region more than the CT urban region. In CT, the correlation increased from (Pearson r = 0.27, p = 2.0e-03 in healthy children to (r = 0.65, p = 1.8e-27) in those with AD (Fig. 3B). A similar pattern was observed in UM, where correlations increased from (r = 0.45, p = 5.6e-14) in healthy children to (r = 0.67, p = 9.8e-36) in children with AD (Fig. 3B). AD-associated microbiome changes in rural umtata and urban cape town We compared children with AD from rural UM versus urban CT to investigate regional differences in disease presentation and skin microbiome composition. Our analyses included 55 children with AD from UM and 44 from CT for the RCLR-based relative abundance comparisons, and 53 children from UM and 43 from CT for the oSCORAD analyses, as oSCORAD scores were not recorded for 2 μm children and 1 CT child. Clinical severity did not differ significantly between regions (t-test, p = 0.584), with similar mean oSCORAD scores in UM (40.5) and CT [30] (Fig. S2A–B). Skin microbiome alpha diversity among healthy children was significantly higher in Umtata compared to Cape Town by Faith PD (Mann-Whitney U-test; p = 0.004, U = 305) (Fig. 4A) and Shannon diversity (p = 0.0049, U = 525), consistent with reports that rural environments harbor greater microbial diversity. To investigate regional differences while controlling for the influence of higher disease severity samples, we performed alpha and beta diversity analyses on two subsets of the data: (i) all samples with oSCORAD values (Fig. S5) and (ii) samples restricted to children with moderate AD (oSCORAD 15–40; Fig. 4B-E). This allowed us to compare regions under similar clinical severity while also ensuring more balanced sample sizes in the moderate-only analysis. Within the alpha diversity analyses, we observed a significant difference between H and ADL only in UM based on Faith PD (p = 0.047, U = 300), with Shannon diversity showing a similar (but non-significant) directional pattern (Fig. S5A–B). No significant difference was observed in alpha diversity from CT groups by either metric, suggesting less contribution from within-sample diversity (Fig. S5C-D). AD-associated microbiome changes in rural Umtata and urban Cape Town. A Among healthy children, Faith PD is significantly higher in Umtata (n = 22) compared to Cape Town (n = 22) (p = 0.0012, U = 104). B–C RPCA beta diversity among children with oSCORAD < 40 shows stronger microbial separation across groups in Umtata (p = 0.001, F = 22.03) than in CT (p = 0.004, F = 4.77). Ellipses denote 90% confidence intervals. D–E box plots (left) and scatterplots (right) display Robust Center Log Ratio (RCLR)-transformed relative abundances of differentially abundant ASVs in urban CT (D) and rural Umtata (E) skin samples. Panels include Streptococcus ASVs, Staphylococcus ASV, and other taxa (Micrococcus, Veillonella_A). Box plots compare abundance across healthy (H), non-lesional AD (ADNL), and lesional AD (ADL) groups. Box plots show medians and interquartile ranges. FDR-corrected p-values from the Mann-Whitney U-test are shown for pairwise groups. Scatterplots show correlations between ranked relative abundance and ranked oSCORAD scores per sample. Spearman correlation coefficients (ρ) and p-values are shown in each scatterplot; grey shading denotes non-significance (p > 0.05) Principal coordinate analysis of robust Aitchison PCA (rPCA) further revealed significant differences in beta diversity between AD and healthy controls in both regions. In UM (Fig. 4B), significant separation was observed among all three skin types including H vs. ADL (PERMANOVA p = 0.001, F = 19.7), H vs. ADNL (p = 0.002, F = 5.8), and ADNL vs. ADL (p = 0.002, F = 6.5). In CT (Fig. 4C), the separation was significant only between H vs. ADL (p = 0.005, F = 4.6), but not between ADNL vs. ADL (p = 0.163, F = 1.9) or H vs. ADNL (p = 0.349, F = 1.1). These patterns were consistent in the analyses which included all oSCORAD measurements (Fig. S5E and F). Across both analyses and both regions, ADNL samples are consistently clustered between H and ADL samples, supporting an intermediate microbiome phenotype. Differential abundance analysis performed with ANCOM-BC2 [31] comparing ADL versus H skin samples revealed substantially more differentially abundant features in UM (Fig. 4G) than in CT (Fig. 4H), with healthy skin samples from CT notably lacking enriched taxa. Based on the high-ranked features from our random forest analysis that were also differentially abundant, we further compared differences in RCLR-transformed relative abundance between skin groups and correlation with oSCORAD. Streptococcus ASV-1 and Streptococcus ASV-2 exhibited significantly higher RCLR-transformed abundances in both ADNL and ADL skin relative to H skin in both UM and CT (Fig. 4D–E). Interestingly, neither ASV significantly correlated with oSCORAD scores, suggesting that Streptococcus enrichment could be a hallmark of AD skin independent of disease severity. In UM, Staphylococcus ASV-1 was also considerably elevated in AD skin compared to healthy controls and correlated positively with oSCORAD (p = 0.26, 2.9e-02). In the CT cohort, we did not see this trend with Staphylococcus (Fig. 4D–E). Additionally, in UM only, Micrococcus ASV-1 was significantly reduced in both ADL and ADNL skin, but higher in H skin. This finding was also consistent with the differential abundance analysis, which showed that Micrococcus ASV-1 was enriched in H samples (Fig. S5G). To explore region-specific differences in nasal microbiota and their association with disease severity, we also assessed ASVs that were differentially abundant in the nares of children with and without AD in Cape Town and Umtata (Fig. S4). In Umtata, Streptococcus ASV-1 abundance was significantly reduced in the nares of children with AD versus healthy children (Mann–Whitney p = 2.4e − 03). In Cape Town, Haemophilus D 734,546 ASV-1 abundance in the nares was significantly higher in AD severity (Mann–Whitney p = 3.7e-02). Discussion Most microbiome research focused on AD has been conducted in a limited number of high-income Western countries [32]. This geographical skew raises questions about the broader applicability of current microbiome-derived insights in the global population, especially regarding health disparities and differing environmental exposures in the Global South [33, 34]. This study helps address a gap by reporting observational findings on the skin and nasal microbiomes of South African children, stratified by rural and urban location. Overall, differences in the microbiome between children with AD and healthy controls were distinct in each geographic region, with the rural cohort exhibiting greater microbiome alteration patterns than the urban cohort. Machine learning has previously been used to identify disease associations from microbial signatures [24], including in understanding atopic dermatitis in children [25]. One of the interesting aspects of this study was our RF model trained solely on nasal samples, which predicted AD status with moderate accuracy (AUC = 0.65 ± 0.06), and models using skin microbiomes could distinguish lesional and non-lesional skin from healthy skin. These findings suggest that skin and nasal microbiome profiles, even from short-read 16S data of a V4-V5 amplified region, can capture disease-related patterns in pediatric AD. While the reported AUCs are likely insufficient for diagnostic screening, these results highlight the potential of microbiome-based classification approaches. Integrating higher-resolution multi-omics data in future studies could enhance early disease prediction, particularly through deeper investigation of non-lesional (clinically healthy-appearing) skin in individuals with AD. A notable observation was the significant similarity between skin and nasal microbiomes in AD patients, supporting the hypothesis that bacterial exchange occurs between these sites. This aligns with the idea that the nasal cavity may act as a reservoir for bacteria capable of exacerbating skin inflammation, such as S. aureus. However, in our analyses, strain overlap and correlation should not be extrapolated to strain identity, as we can only resolve to the Genus level with 16S data. In our cohort, Streptococcus ASVs were consistently enriched in AD skin, yet showed no correlation with oSCORAD scores. Whether Streptococcus is more associated with disease presence than progression, and whether it can facilitate colonization by Staphylococcus species such as S. aureus, warrants further investigation. Assessing the nasal microbiome is vital for evaluating the risk of cutaneous colonization by potentially pathogenic species. Additional research is needed to clarify the mechanisms and directionality of this exchange, as well as to determine whether the nasal microbiome plays a causal role in the development of cutaneous dysbiosis. We acknowledge several methodological limitations in our study design. Participants were enrolled at a single time point. While those requiring topical or systemic antibiotics for acute AD flares were excluded, we did not record the time elapsed since their last use of antibiotics. All participants with AD were receiving standard-of-care treatment, including soap substitutes, emollients, and topical corticosteroids at the time of sampling. Given the young age of our cohort and disease severity, withholding treatment was ethically unfeasible; these therapeutic interventions were necessary to maintain participant welfare and comfort. However, prior research suggests that the dysbiotic microbial signatures in AD persist despite immunomodulatory therapy; notably, studies of dupilumab treatment have demonstrated that S. aureus dysbiosis persists even with blockade of the IL-4/IL-13 pathway [35]. While topical corticosteroids alleviate inflammation, our findings suggest that key microbiome differences between AD and healthy skin persist, indicating that these microbial alterations may represent a fundamental feature of AD pathophysiology rather than merely a consequence of acute inflammation. Our study was limited in resources to only 16S sequencing rather than shotgun metagenomics. 16S rRNA gene sequencing restricts taxonomic resolution to the genus level, thereby limiting insights into species and strain levels. Our study utilized the 16S V4–V5 region, which underrepresents Cutibacterium, an abundant skin bacterial genus that is more accurately captured by V1–V3 primers [36]. The V4–V5 region (~ 400 bp) is longer than V4 alone (~ 250 bp) and should provide more phylogenetic coverage in general, which is why it has been used in many microbiome studies [37, 38]. In this study, the V4–V5 region successfully captured Streptococcus, Staphylococcus, and numerous other known skin microbiome taxa. However, to our knowledge, there are currently no published studies that have systematically evaluated whether the V4-V5 region improves the identification of Staphylococcus on the skin compared to V4 alone for skin microbiome studies. We also declare that lesional body sites were not recorded for each sample. However, sampling was conducted from typical pediatric AD predilection sites (e.g., cheeks, forearms) using standardized protocols across both locations. Additionally, all non-lesional samples from children with AD were collected from the mid-back. This sampling design introduces a bias that prevents the independent assessment of disease effects and natural skin biogeography. This limitation is particularly relevant given that the skin microbiome differs markedly across body sites in both healthy individuals and those with AD [39, 40] and should be noted as a caveat. African skin is compositionally distinct from that of Western populations, underscoring the need for region-specific studies to understand better its roles in health and disease [41, 42]. The skin microbiome is further influenced by urbanization [30], with rural and less urbanized environments typically associated with greater bacterial diversity [43–46], which may be protective in early life. Based on a 2019 study from the South African Food Allergy cohort, which sampled 535 urban children and 347 rural children in South Africa, 25.6% of urban children had AD compared to 2.0% of rural children [47]. These epidemiological differences underscore the importance of examining AD microbiomes across diverse environmental contexts, including urbanization. While our study design, which involves sampling at one urban and one rural site, limits broader generalizability, our findings expand the available data on pediatric AD microbiomes in South Africa and provide evidence that the environmental context contributes to the microbial dysbiosis driven by the disease. Future studies should collect improved metadata on the environmental setting to better distinguish between biogeographic and disease-driven effects. While our results contrast with the traditional “hygiene hypothesis”, which posits that rural environments protect against allergic diseases through greater microbial exposure, children in rural UM exhibited stronger microbial dysbiosis than those in urban CT, even after controlling for disease severity. This pattern may reflect the higher baseline skin microbiome diversity observed among healthy UM children compared to their CT counterparts, as well as differences in healthcare access, treatment practices, and antibiotic exposure between sites. Lesion body sites were not recorded, which may further contribute to variability in the results. These findings should therefore be interpreted within the broader context of environmental, clinical, and immunologic differences between regions. Cape Town represents an urban, coastal, tertiary-care recruitment setting. In contrast, Umtata is a rural, inland region with distinct environmental exposures and more limited healthcare access. Together, these contextual factors likely also underlie the greater microbiome alterations observed in the Umtata cohort. Finally, a study in a subset of this cohort [48] found that rural children, regardless of AD status, displayed heightened innate immune activation and AD-associated gene expression, likely reflecting environmental exposures such as contact with farm animals during pregnancy. While AD was associated with downregulated lymphocyte and innate signaling pathways, the environment exerted a substantial influence on immune transcriptional profiles [49]. Rural children exhibited a marked upregulation of chemokines, cytokines, GPCRs, and regulatory pathways, including IL-10, a key anti-inflammatory cytokine that suppresses antigen-presenting cells and lymphocyte effector functions. These environmental, immunologic, and healthcare-access differences may help explain the greater microbiome alterations observed in rural children with AD. Methods Clinical study design and sample collection We collected a total of 502 skin and nasal samples from 197 South African children between February 2015 and May 2016. The complete sampled cohort comprised 110 children clinically diagnosed with atopic dermatitis (AD) and 87 healthy controls, aged 9–37 months. Our clinical collection protocol prioritized obtaining high-quality swabs from all accessible body sites where a child presented with adequate lesional or non-lesional areas for sampling, which resulted in some children contributing multiple samples across different body sites or sample types. To ensure statistical independence and avoid pseudoreplication, we subsetted the data to include only one sample per child per body site, resulting in a final analytical dataset of 462 samples from 183 unique children. This comprised 99 AD lesional skin samples, 99 AD nonlesional skin samples, 96 AD anterior nares samples (nasal samples were not collected from 3 children with AD), 84 control nonlesional skin samples, and 84 control anterior nares samples. All results presented in this manuscript reflect analyses performed on this deduplicated dataset. For full transparency and reproducibility, we provide complete metadata for both the original 502 samples (16S_AD_South-Africa_metadata_full.tsv) and the 462 samples included in final analyses (16S_AD_South-Africa_metadata_subset.tsv), along with a Python notebook documenting sample count calculations (Fig_1_metadata-counts.ipynb). Demographic and clinical characteristics of study participants are summarized in Table 1. AD diagnosis followed the UK Working Party’s criteria [50] as assessed by trained clinicians. Disease severity was evaluated using the objective SCORAD (oSCORAD) index, which ranges from 0 to 83 and is based on the extent of affected body surface area and the intensity of six clinical signs (erythema, edema/papulation, oozing/crusts, excoriation, lichenification, and dryness), each scored from 0 to 3. The final score was computed using the formula: oSCORAD = A/5 + 7B/2 [51]. Children with chronic illnesses, immunomodulatory treatments, or current antibiotic use at the time of sampling were excluded; however, the time since their last antibiotic use was not recorded. All children with AD were receiving standard-of-care therapy, including soap substitutes, emollients, and topical corticosteroids at the time of enrollment, as withholding treatment would have been unethical. Healthy, non-atopic, age-matched, and sex-matched participants were eligible as controls if they had no personal history of atopy and were non-food sensitized by skin prick testing. Both AD and healthy participants were excluded if they had any other skin or systemic disease. Participants were recruited from two socio-environmentally distinct locations: Cape Town (CT; ~5 million population), a highly urbanized setting, and Umtata (UM; ~200,000 population), a rural region. Samples from urban AD cases were collected at Red Cross Children’s Hospital in CT, and healthy controls from local preschools. Rural samples were collected at Nelson Mandela Academic Hospital in UM, primarily from the Mqanduli district. All participants were of Black (AmaXhosa ethno-linguistic) descent. Written informed consent was obtained from parents or legal guardians. Skin and nasal swabs were collected using standardized protocols [50, 51] and performed consistently across both sites. For AD children, lesional skin swabs were taken from commonly affected sites (e.g., cheeks, forearms), and non-lesional swabs from the mid-back. Specific lesional body site locations were not recorded. Healthy controls were swabbed on the mid-back for consistency. Nasal swabs were taken from the anterior nares. All samples were stored at − 80 °C in skim milk-tryptone-glycerol-glucose buffer until processing. Negative controls included swabs, Milli-Q water, pure negatives, and STGG media controls. These were processed through the same steps as all other samples. DNA extraction, 16S rRNA gene sequencing, and sequence data processing DNA was extracted from skin and nasal swabs using the DSP Virus/Pathogen Mini Kit® (Qiagen, Germany) with a 70 µL elution volume at the University of Cape Town. The V4–V5 region of the 16S rRNA gene was amplified using previously described primers 515 F (TATGGTAATTGTGTGYCAGCMGCCGCGGTAA) and 926R (CGGCATACGAGATAGTCAGCCAGGGCCGYCAATTYMTTTRAGTTT) [52–54]. While this primer pair underrepresents Cutibacterium relative to V1–V3 primers [36], it provides broader coverage of Staphylococcus and Streptococcus sequencing depth and is widely used in microbiome studies [55]. Library preparation and sequencing were conducted at the J. Craig Venter Institute in San Diego, CA. Amplicons were barcoded, pooled in batches of 300, and sequenced on an Illumina MiSeq v3 (300-cycle) run, generating ~ 13 Gbp per run. Raw data were processed via the Qiita platform [56] using the standardized 16S V4-V5 workflow. Sequences were demultiplexed, PHRED [57] quality-filtered, and denoised using Deblur [58] in QIIME 2 (v2024.2.0). ASVs were assigned taxonomy using the Greengenes 2 database (v2022.10) [59], and feature tables were generated at both ASV and genus levels. To reduce sparsity from extremely minor and possibly spurious features, a 1% sample depth threshold was applied to filtered tables. We analyzed the rarefaction curve with QIIME2 to determine a rarefaction depth. We rarefied samples to a depth of 2,000 reads for alpha diversity analysis, excluding samples below this threshold. Beta diversity analyses were performed on unrarefied, prevalence-filtered count tables. Random forest classification RF classifiers were implemented using the scikit-learn Python package (v1.4.2) to evaluate the predictive power of 16S rRNA ASV-level features (n = 2283) and their observed abundances across multiple classification tasks. Models were trained to distinguish (1) skin and nasal samples, (2) ADL, ADNL, and H skin samples in pairwise comparisons (ADL vs. H, ADNL vs. H, ADL vs. ADNL), and (3) AD vs. H status in nasal samples. For random forest classification, we used non-prevalence filtered, non-rarefied feature tables to maximize information available to the model, as this machine learning approach can inherently handle sparse data and identify informative features. To account for potential confounding factors within the RF analysis, we included metadata variables, including age_months, sex, and enrollment_season, as features within the model. To prevent overfitting and account for repeated measures, we ensured all samples from a participant were assigned exclusively to either the training (80%) or testing (20%) set in each fold (GroupKFold, n_splits = 5). RF models with 1,000 estimators were trained for each classification task, and performance was evaluated using the area under the ROC curve. Feature importance was measured by the mean decrease in Gini impurity across folds, and the top 10 ASVs were retained to identify key bacterial drivers of predictive performance. To assess the directionality of each feature, we performed log-fold-change analysis, with positive values (+) indicating higher abundance in the first comparison group and negative values (−) indicating higher abundance in the second comparison group. To explore geographic variation in classification performance, the above analyses were repeated independently for CT and UM cohorts using the same classification framework (Fig. S4). ROC curves shaded regions denoting ± 1 SD around the mean true positive rate. This comprehensive framework enabled us to assess cross-site consistency and regional specificity in microbiome-based disease status and anatomical site classification. Microbiome diversity analyses Alpha diversity was assessed using the Faith PD, an evolutionarily informed metric that incorporates a phylogenetic tree [60]. The phylogenetic tree was constructed through the 16S ASVs through multi-sequence alignment with MAFFT, and the tree was made with FastTree. ASV count tables were processed using the BIOM (v.2.1.15) package via Python (v.3.10.14). We conducted all alpha diversity analyses with rarefied tables at a depth of 2000 reads as noted above. This analysis included only AD samples with oSCORAD scores of 15 to 40 and healthy controls to minimize potential confounding from extreme disease severity. Faith PD values were computed for each sample using the skbio.diversity.alpha_diversity function. Statistical significance was determined using the non-parametric Mann–Whitney U-test for all pairwise comparisons. All p-values were FDR-adjusted using BH. The same analysis with all samples (including oSCORAD >40) is in the supplementary. Beta diversity was analyzed using RPCA [61], a compositionality-aware ordination technique within gemelli (v.0.0.12). All beta diversity analyses were conducted with non-rarefied tables. Before RPCA, samples were filtered to exclude nasal swabs, and the feature table was converted back into BIOM format to ensure compatibility with the gemelli.rpca function. Ordination was performed on this filtered table, and sample coordinates from the first two RPCA axes were used to visualize group separations. Statistical differences in beta diversity were tested using PERMANOVA on the RPCA-based distance matrix, with significance evaluated between pairwise combinations of skin sample groups. RCLR-transformed relative abundance analysis was performed using feature tables rarefied to 350 reads to maximize sample retention while focusing on major taxa. Differential abundance analysis was conducted using ANCOM-BC2 [31] with non-rarefied data. A formula of “case_type + age_months + sex + enrollment_season” was used to account for potential confounding. Within both alpha and beta diversity analyses, we evaluated potential confounding factors that could influence interpretation. Specifically, we assessed the metadata variables age_months, sex, and enrollment_season, as detailed in Confounding_Assessment.ipynb. No evidence of confounding was identified for the alpha diversity results. In contrast, the beta diversity analysis showed that age_months and enrollment_season differed significantly between regions and could confound group-level differences. Therefore, for beta diversity we used PERMANOVA models that included these variables as covariates (i.e., distance_matrix ~ group + age_months + enrollment_season) to adjust for their effects while testing for differences by disease status and region. All analyses were conducted in Python (v3.10.14), except for differential abundance testing (ANCOM-BC2), which was performed in R (v.4.3.3). Supplementary Information Supplementary Material 1. Study Metadata. Supplementary Material 2: Figure S1. RF classification performance using 16S ASV profiles distinguishes skin and nares microbiomes between locations. (A) Rural UM. (B) Urban CT. Each panel shows the ROC curves and AUC values. Left panels compare skin versus nares samples in both UM (AUC=0.93 ± 0.01) and CT (AUC=0.97 ± 0.02). The middle panels show classification of AD status based on skin samples, with separate models for each pairwise comparison. In UM, skin-based models show strong performance (AUCs=0.75–0.87), as did CT skin-based models (AUCs=0.70–0.85). Right panels assess AD status prediction from nares samples in both UM (AUC=0.61 ± 0.10) and CT (AUC=0.83 ± 0.07). Shaded regions represent the standard deviation across cross-validation folds. Figure S2. Comparison of model performance across binary classification tasks. Heatmap showing the mean area under the receiver operating characteristic curve (AUC-ROC) for four supervised machine learning classifiers, Random Forest, XGBoost, LightGBM, and Support Vector Machine (SVM), across five binary classification tasks. Each cell represents the average AUC-ROC value, with darker blue indicating stronger model performance. Figure S3. Distribution of oSCORAD severity scores of AD cases in Umtata and Cape Town. (A–B) oSCORAD severity scores show more severe AD cases in Umtata (n=54) compared to CT (n=43); dashed lines indicate median scores. The number of AD children from each geographic location was 55 in UM and 44 in CT. However, two children from Umtata and one child from Cape Town did not have a recorded oSCORAD value. Difference in severity distribution was not significant between UM and CT (T-statistic=-0.550, p= 0.584). Figure S4. Healthy skin microbial alpha diversity differs by geography. Boxplots show Shannon diversity among healthy skin samples collected from children in Umtata (rural) and Cape Town (urban). Each point represents one individual. Horizontal lines indicate medians and interquartile ranges. A two-sided Mann–Whitney U test revealed higher microbial diversity in Umtata compared to Cape Town (U=525, p=0.0049). Figure S5. Skin microbiome differences by disease status and geographic region (all oSCORAD values included). (A–D) Alpha diversity of skin microbiomes from healthy (H), acne non-lesional (ADNL), and acne lesional (ADL) samples in rural Umtata (UM; panels A–B) and urban Cape Town (CT; panels C–D). Faith’s Phylogenetic Diversity (PD) and Shannon diversity metrics are shown. Boxes represent the median and interquartile range (IQR), whiskers extend to 1.5× IQR, and each point represents an individual sample. Pairwise comparisons were performed using Mann–Whitney U tests with FDR correction. (E–F) Beta diversity of skin samples visualized using robust Aitchison principal component analysis (RPCA) for UM (E) and CT (F). Ellipses denote 95% confidence intervals for each group. Pairwise differences between groups were assessed using PERMANOVA, with F and p-values shown. (G–H) Differential abundance analysis comparing ADL versus H skin samples from UM (G) and CT (H). Volcano plots display log₂ fold-changes and –log₁₀ p-values from ANCOM-BC2, adjusted for age, sex, and enrollment season. Colored points represent ASVs passing both fold-change and significance thresholds; labeled features highlight the top differentially abundant taxa. Figure S6. Region-specific associations between nasal bacterial taxa and AD severity. Relative abundance RCLR-transformed of ASVs in the nares of children from urban CT (A) and rural UM (B), shown by disease status (box plots) and against oSCORAD scores (scatter plots). Box plots show medians and IQR. Spearman correlation coefficients (ρ) and p-values are shown in each scatterplot; grey shading denotes non-significance (p > 0.05). Acknowledgements We are grateful to the clinicians at Red Cross Children’s Hospital, Cape Town, and the Nelson Mandela Academic Hospital, Umtata, for collecting the samples for this project. Authors’ contributions Y.C. led conceptualization, data curation, formal analysis, investigation, methodology, software, validation, visualization, and preparation of all figures, writing of the original draft, and also led review and editing. A.N. contributed to data curation, formal analysis, investigation, visualization, and writing (original draft and review/editing). N.L. contributed to methodology, validation, and manuscript review/editing. G.N. contributed to conceptualization, funding acquisition, investigation, methodology, and review/editing. D.K. contributed to formal analysis. B.D.P. contributed to data curation, validation, and review/editing. T.N. contributed to conceptualization.M.M. contributed to project administration and manuscript review/editing. E.S. contributed to project administration and manuscript review/editing.C.H. contributed to funding acquisition, methodology, and project administration. M.L. contributed to manuscript review/editing. A.M. contributed to methodology and project administration. G.H. contributed to manuscript review/editing. C.C. contributed to manuscript review/editing. R.K. contributed to data curation, formal analysis, investigation, methodology, supervision, and review/editing. R.G. contributed to conceptualization, investigation, supervision, and review/editing. C.D. led conceptualization, data curation, investigation, methodology, project administration, funding acquisition, and supervision, and contributed to formal analysis and review/editing. F.D. led conceptualization, investigation, methodology, project administration, funding acquisition, and supervision, and contributed to data curation and review/editing.All authors reviewed the manuscript. Funding This research was funded by the National Research Foundation of South Africa (UID: 150539). Sequencing was supported by NIH 1U54GH009824. Data availability Sequencing 16S rRNA gene sequencing data is deposited in the NCBI Sequence Read Archive (SRA) under BioProject ID [PRJNA1269634](https://www.ncbi.nlm.nih.gov/bioproject/?term=PRJNA1269634). The data processing can be found on QIITA via [study ID 15564](https:/qiita.ucsd.edu/study/description/15564). Code used in our analyses, along with the manuscript figures, can be found at our GitHub repository: https://github.com/yangchen2/16S_AD_South-Africahttps://github.com/yangchen2/16S/_AD/_South-Africa. Declarations Ethics approval and consent to participate The Human Research and Ethics Committee of the Faculty of Health Science, University of Cape Town (HREC/REF: 451/2014) approved the parent study, and additional ethical approval for the bacteriome study, Human Research and Ethics Committee (HREC/REC: 668/2020), was obtained. Written informed consent was obtained from all parents or legal guardians of children included in this study. All research activities were in compliance with the ethical principles under the World Medical Association Declaration of Helsinki. Consent for publication by study participants is not applicable as no identifying information is included. Competing interests R.L.G. serves as a co-founder, scientific advisor, consultant, and equity holder in MatriSys Biosciences, and holds a position as a consultant, receiving income, and equity holder in Sente Inc. R.K. is a scientific advisory board member and consultant for BiomeSense, Inc., holding equity and receiving income. He is a member of the scientific advisory board and has equity in GenCirq. He has equity in and acts as a consultant for Cybele. He is a co-founder of Biota, Inc., and has equity. He is a co-founder of Micronoma, holds equity, and serves on the scientific advisory board. He is a board member of Microbiome Vault, Inc. He is a N=1 IBS advisory board member and receives income. He is a Senior Visiting Fellow of the HKUST Jockey Club Institute for Advanced Study. The terms of these arrangements have been reviewed and approved by the University of California, San Diego, in accordance with its conflict-of-interest policies. C.C. is the founder of Phiome and receives income from Dr. Armpit. He serves as a scientific advisor to NAOS, MyMicrobiome, and Parallel Health, Inc. C.L.D. is a co-founder, president, chief strategic officer, and equity holder of Biorentis, Inc. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1. Study Metadata. Supplementary Material 2: Figure S1. RF classification performance using 16S ASV profiles distinguishes skin and nares microbiomes between locations. (A) Rural UM. (B) Urban CT. Each panel shows the ROC curves and AUC values. Left panels compare skin versus nares samples in both UM (AUC=0.93 ± 0.01) and CT (AUC=0.97 ± 0.02). The middle panels show classification of AD status based on skin samples, with separate models for each pairwise comparison. In UM, skin-based models show strong performance (AUCs=0.75–0.87), as did CT skin-based models (AUCs=0.70–0.85). Right panels assess AD status prediction from nares samples in both UM (AUC=0.61 ± 0.10) and CT (AUC=0.83 ± 0.07). Shaded regions represent the standard deviation across cross-validation folds. Figure S2. Comparison of model performance across binary classification tasks. Heatmap showing the mean area under the receiver operating characteristic curve (AUC-ROC) for four supervised machine learning classifiers, Random Forest, XGBoost, LightGBM, and Support Vector Machine (SVM), across five binary classification tasks. Each cell represents the average AUC-ROC value, with darker blue indicating stronger model performance. Figure S3. Distribution of oSCORAD severity scores of AD cases in Umtata and Cape Town. (A–B) oSCORAD severity scores show more severe AD cases in Umtata (n=54) compared to CT (n=43); dashed lines indicate median scores. The number of AD children from each geographic location was 55 in UM and 44 in CT. However, two children from Umtata and one child from Cape Town did not have a recorded oSCORAD value. Difference in severity distribution was not significant between UM and CT (T-statistic=-0.550, p= 0.584). Figure S4. Healthy skin microbial alpha diversity differs by geography. Boxplots show Shannon diversity among healthy skin samples collected from children in Umtata (rural) and Cape Town (urban). Each point represents one individual. Horizontal lines indicate medians and interquartile ranges. A two-sided Mann–Whitney U test revealed higher microbial diversity in Umtata compared to Cape Town (U=525, p=0.0049). Figure S5. Skin microbiome differences by disease status and geographic region (all oSCORAD values included). (A–D) Alpha diversity of skin microbiomes from healthy (H), acne non-lesional (ADNL), and acne lesional (ADL) samples in rural Umtata (UM; panels A–B) and urban Cape Town (CT; panels C–D). Faith’s Phylogenetic Diversity (PD) and Shannon diversity metrics are shown. Boxes represent the median and interquartile range (IQR), whiskers extend to 1.5× IQR, and each point represents an individual sample. Pairwise comparisons were performed using Mann–Whitney U tests with FDR correction. (E–F) Beta diversity of skin samples visualized using robust Aitchison principal component analysis (RPCA) for UM (E) and CT (F). Ellipses denote 95% confidence intervals for each group. Pairwise differences between groups were assessed using PERMANOVA, with F and p-values shown. (G–H) Differential abundance analysis comparing ADL versus H skin samples from UM (G) and CT (H). Volcano plots display log₂ fold-changes and –log₁₀ p-values from ANCOM-BC2, adjusted for age, sex, and enrollment season. Colored points represent ASVs passing both fold-change and significance thresholds; labeled features highlight the top differentially abundant taxa. Figure S6. Region-specific associations between nasal bacterial taxa and AD severity. Relative abundance RCLR-transformed of ASVs in the nares of children from urban CT (A) and rural UM (B), shown by disease status (box plots) and against oSCORAD scores (scatter plots). Box plots show medians and IQR. Spearman correlation coefficients (ρ) and p-values are shown in each scatterplot; grey shading denotes non-significance (p > 0.05). Data Availability Statement Sequencing 16S rRNA gene sequencing data is deposited in the NCBI Sequence Read Archive (SRA) under BioProject ID [PRJNA1269634](https://www.ncbi.nlm.nih.gov/bioproject/?term=PRJNA1269634). The data processing can be found on QIITA via [study ID 15564](https:/qiita.ucsd.edu/study/description/15564). Code used in our analyses, along with the manuscript figures, can be found at our GitHub repository: https://github.com/yangchen2/16S_AD_South-Africahttps://github.com/yangchen2/16S/_AD/_South-Africa.