Background

Childhood undernutrition is a significant global public health issue contributing to approximately 45\(\%\) of deaths in children under five years of age [1]. Undernutrition not only poses a high risk of mortality but also causes growth stunting, immune dysfunction, neurocognitive deficits, and endocrine system disorders [2, 3]. Although providing nutrient supplementation and antibiotics is the current therapeutic approach for undernutrition [4, 5], the long-term use of antibiotics is controversial, and alternative therapeutic strategies are necessary [6, 7].

In recent years, the gut microbiome (GM) has emerged as a promising target for the treatment of undernutrition due to its crucial role in nutrient metabolism and immune regulation. Previous studies have identified the association between GM composition and undernutrition. By studying a large child cohort in Bangladesh for two years, Subramanian et al. found the persistent immature GM in undernutrition (UN) children [7]; and the immature GM led to the mouse growth deficit after transferring to a mouse [6, 8]. Yan et al. reported that GM participated in the secretion of insulin-like growth factor 1 (IGF-1), promoting bone formation and growth [9]. Also, acetate, a short-chain fatty acid, released by the microbiome fermentation, affects the body’s adiposity [10]. However, these investigations have largely been conducted at the species or functional level, and the specific contributions of individual GM strains remain unclear.

To address this knowledge gap, we present a novel strategy to detect GM compositions at the strain level in undernourished children. Our approach leverages metagenomic sequencing and machine learning algorithms to identify specific GM strains associated with undernutrition. By examining the GM at the strain level, we aim to gain a more comprehensive understanding of the complex relationship between GM and undernutrition, potentially identifying new avenues for developing GM-based therapies for undernutrition.

Methods

Data preparation

The metagenomic sequencing data from 61 undernutrition patients (UN group) and 36 healthy children (HC group) was obtained from our previous study (the NCBI Sequence Read Archive Database under the accession number PRJNA543967) [11]. DNA libraries were sequenced with the 150 bp paired-end mode on the HiSeq platform (Illumina, San Diego, CA, United States). The concentrations of hemoglobin, albumin, total protein, white blood cells, and lymphocytes were measured by the automatic blood analyzer (Beckman Coulter AU5800, Brea, CA, USA) using peripheral blood. All recruited children were under three years old. The UN children had moderate or severe undernutrition, which was defined as having a weight-for-age z-score two standard deviations below the WHO reference value. The HC children were selected from those who passed physical examinations with no diarrhea in the past two weeks. Moreover, our previous study excluded participants who had been exposed to antibiotics, probiotics, or proton pump inhibitors the month before fecal sample collection, had a known history of allergies or hereditary diseases, suffered from metabolic or autoimmune diseases, or had parasitic eggs in their stool. The sequencing reads were filtered with the following criteria: 1) The raw reads containing more than 10 low-quality (<Q20) bases or 15 bases of adapter sequences were filtered out; 2) The raw reads which can be aligned to human genome hg19 by BWA-MEM were filtered out. After the filtration, the remaining reads were applied for the downstream analysis.

Species taxonomic annotation

To obtain the taxonomic information for the samples, we applied the MetaPhlan3 software to align the clean data to the marker gene database (Version: mpa_v30_CHO- COPhlAn_201901) [12] and combine the taxonomical annotation files. We removed the species absent in 70\(\%\) of the samples. At last, a total of 453 species remained for further analysis.

Strain-level identification

We applied PStrain [13] to identify strains and infer strain abundances for each sample. Strains with a sequence similarity of more than 90\(\%\) clustered together, and the identified strains were named by their cluster name. PStrain identified a total of 1,932 strain clusters. To annotate these strain clusters, we used PStrain-tracer [14] to obtain NCBI reference sequences for the species the strains belong to and construct the phylogenetic trees for the strain clusters and NCBI reference sequences [15]. Then, we used an in-house Python script to annotate the strain clusters with the closest NCBI known strains in the phylogenetic tree. With the strain profiling and identified strains, we completed the subsequential statistical analysis.

Bacterial co-occurrence network and inner bacterial interactions

Based on the taxonomical profiling at the strain level, we calculated the Spearman correlation coefficient among strains by using the R software package “psych” [16]. The relations whose Spearman correlation coefficient <-0.6 or > 0.6 (adjusted P < 0.05) were kept as the edges for the co-occurrence network. In addition, we calculated the network topology information such as average degree, average path length, and cluster coefficient for each group. At last, the bacterial co-occurrence networks were visualized by Gephi (Version 0.9.2), with the nodes representing strains and edges representing the correlation between strains [17]. The species are colored by their corresponding phylum, and the red and pink edges stand for the positive and negative correlations, respectively. We further measured the network topological structure with the R package “igraph” on the constructed co-occurrence networks [18].

Construction of Random Forest classifiers

Applying the Random Forest in machine learning (package “random-Forest” in R), we constructed classification models for the prediction of the UN and HC participants at the species and strain levels, respectively [19]. For the construction of the model, we randomly divided the total of 97 samples into five groups, with three groups containing 19 samples each and the other two groups containing 20 samples each. Then, we used the five-fold cross-validation method during the model construction. Four groups were selected as the training dataset, and one group was taken as the validation dataset randomly. Third, the Random Forest model was trained with the training dataset, and the performance of the constructed models was assessed with the validation dataset. At last, we plotted the Receiver Operating Characteristic (ROC) curves and calculated the AUC values to visualize the performance of the classifier (package “pROC” in R). In addition, the species and strains were recognized as candidate biomarkers for the UN prediction according to their Gini values and the optimal variation numbers in the Random Forest models.

Statistical analysis

We evaluated the bacterial diversity using Shannon and Chao1 index at species and strain levels with R package “vegan” [20]. Then, we adopted the Wilcoxon rank sum test to explore the differentially enriched species and strains between different groups. The results were plotted with package “ggplot2” in R. In addition, we used Benjamini and Hochberg method (“p.adjust” in R) for the adjustment of the multiple Wilcoxon rank sum tests and Spearman correlation analysis (FDR<0.05) [21]. The Spearman correlation analysis results were visualized by the R package “pheatmap”. We denoted the nutrition-related clinical factors as nutrition indicators. Then we calculated the pairwise Pearson correlation between nutrition indicators, and the contribution of GM, short-chain fatty acids (SCFAs), and Weight to Age by the Mantel test [22]. The calculation and visualization were performed by the R package “linkET” [23].

Results

UN and HC children have similar strain diversity

To obtain the subtle bacterial signals related to the UN pathogenesis, we applied the PStrain software and detected the GM differences between the UN and HC children at the strain level [13]. The PCoA, Shannon and Chao1 diversity exhibited the overall GM differences between the UN and HC groups (Fig. 1a and b). The bacterial diversity in UN children was lower than that of HC children, although only Chao1 showed statistical differences (P = 0.01), while Shannon did not (P = 0.41, Fig. 1b). Then, we performed the abundance comparisons for the bacteria between the two groups at the species and strain levels, respectively. At the species level, we found six of the top ten abundant species were differentially enriched between UN and HC children (Fig. 1c). Among them, Bacteroides fragilis (Padj = 0.005), Bacteroides uniformis (Padj < 0.001), Bifidobacterium pseudocatenulatum (Padj < 0.001), Faecalibacterium prausnitzii (Padj < 0.001), and Bacteroides vulgatus (Padj = 0.008) were enriched in the HC group, while Prevotella copri (Padj < 0.001) was enriched in the UN group. Subsequently, we compared the differences of the top ten abundant strains between the UN and HC groups. We discovered that B. fragilis ASM181622v2 (Padj = 0.008) and F. prausnitzii (Padj = 0.012) were enriched in the HC group (Fig. 1d), indicating that the strain-level analysis would provide us with more comprehensive information for UN pathogenesis investigation.

Fig. 1
figure 1

GM features in the UN patients. a PCoA with Bray-Curtis distances among samples. Each spot stands for a sample, and the circles contain samples with a 90\(\%\) confidence interval for the groups. Red and blue colors stand for the UN and HC children, respectively. b Shannon diversity in the UN and HC children. The red and blue violins stand for the distributions of Shannon indices in the UN and HC groups, respectively. c Comparison of the top 5 abundant species between the UN and HC children. Each plot contained a species; the red and blue boxes stand for the UN and HC groups, respectively. d Comparison of the top 10 abundant strains between the UN and HC children. The red and blue columns stand for the UN and HC groups, respectively. *, Padj < 0.05; **, Padj < 0.01; ***, Padj < 0.001

Full size image

Strain compositional differences explain discriminated species in the UN children

With the NR database and PStrain-tracer software, we obtained the strain-level profiling and discovered the specific strain compositional characteristics for the differentially enriched species between the UN and HC children (Fig. 2, Supplementary Figs. 1, 2, 3, 4, and 5, and Supplementary Table 1). First, the strain composition of the same species in the UN patients was completely different from the HC children. For example, P. copri was mainly composed of P. copri ASM1680343v1 and P. copri ASM345920v1 in the HC group, while it was composed of P. copri ASM346549v1 and P. copri ASM347465v1 in the UN group (Fig. 2e). Second, the common strains between the UN and HC groups had different abundances, leading to species-level differences. For example, the enriched B. fragilis ASM181622v2 in the HC group was responsible for the B. fragilis discrimination between the UN and HC groups (Fig. 2a), while B. vulgatus sp. contributed to the B. vulgatus difference between the UN and HC groups (Fig. 2c). Meanwhile, highly abundant B. pseudocatenulatum BPSEU7765_v1, B. pseudocatenulatum ASM168596v1, and B. pseudocatenulatum ASM346751v1 in the HC group laid the basis for the enrichment of B. pseudocatenulatum in the HC group (Fig. 2d). Finally, the strain diversity for the same species might be different between the UN and HC groups. For example, the strain diversity of B. uniformis in the HC group (Si = 13) was higher than that in the UN group (Si=9, Fig. 2b). Based on these findings, we concluded that the strain-level annotation results can not only help us obtain subtle information on the GM compositions but also provide an important basis for further accurate GM intervention in diseases.

Fig. 2
figure 2

Strain components of the differentially enriched species in the UN patients. The plots a, b, c, d, and e exhibited the components and abundances of Bacteroides fragilis, Bacteroides uniformis, Bacteroides vulgatus, Bifidobacterium pseudocatenulatum, and Prevotella copri in the UN and HC groups, respectively. The number of the identified strains for each species was labeled below the plots for each group, and their abundances were represented by the height of the columns

Full size image

The UN children exhibited a more complex but lower efficient GM co-occurrence network

To further detect the hub bacteria that manipulate the GM, we constructed the GM co-occurrence networks for the UN and HC children and compared their topological features with a complex network approach. We discovered that the UN children had a larger network than the HC children, including more nodes and edges (Fig. 3a and b). In addition, the degree distributions of nodes in both UN and HC were confronted with the power-law distribution, implying that they are both scale-free networks (Fig. 3c and d). The GM network in HC children contained 19 hub nodes, including Bacteroides stercoris ASM343802v1, Alistipes putredinis ASM1913202v1, Butyricicoccus pullicaecorum, etc. (Fig. 3a, Supplementary Table 2). In UN children, the GM network contained 21 hub nodes, including A. putredinis, Ruminococcus bromii ASM346616v1, Barnesiella intestinihominis, etc. (Fig. 3b, Supplementary Table 3). The topological structure analysis indicated that the GM network in UN children has a higher clustering coefficient (0.7780 vs. 0.7700) and modularity (0.7780 vs. 0.7638), indicating their more complex network. However, the graph density (Fig. 3h) and the average degree of each bacterium (Fig. 3e) in the UN patients were lower than those in the HC children, and the average path length was longer in the UN children (Fig. 3f), suggesting the connectivity and information transferring efficiency was lower in UN patients than that of HC children. The finding suggested that the UN children had a discriminated GM network, providing us with bacterial candidates for further GM interventions in these patients.

Fig. 3
figure 3

Topological features of the GM co-occurrence networks in the HC and UN patients. Plot a and b describe the GM co-occurrence networks in the HC and UN, respectively. The number of nodes and edges for the networks are labeled under each plot. Each circle stands for a strain, and its diameter represents the abundance of the strain. The lines stand for the correlations among the strains, while the positive correlations were marked with red color. Plot c and d describe the degree distributions of the nodes for the GM networks in the HC and UN groups. Plot e, f, g, h, and i exhibited the average degree, average path length, clustering coefficient, graph density, and modularity in the UN and HC groups, respectively

Full size image

Improved predictive power of UN-risk model at the strain level

With the random forest algorithm, we constructed the UN-risk models at species and strain levels, respectively, and identified microbial markers for the UN patients (Fig. 4). After constructing the UN-risk models, we recognized the top five variables with significant contributions as the potential biomarkers for the UN patient screening. At the species level, B. uniformis, B. vulgatus, and P. copri were the potential biomarkers (Fig. 4a). Whilst, B. fragilis ASM181622v2, Ruthenibacterium lactatiformans, and Clostridium neonatale sp. were the potential biomarkers at the strain level (Fig. 4b). The findings suggested that the potential biomarkers were uncoordinated at species and strain levels. In addition, the UN-risk model constructed from the strain compositions exhibited higher accuracy (area under curve, AUC = 0.833) than that from the specie compositions (AUC = 0.743), implying the higher accuracy UN-risk model constructed with strain profiling.

Fig. 4
figure 4

UN-risk models constructed with GM at the species and strain levels. a The ROC curves and Gini values for the UN-risk model constructed at the species level. B. uniformis, B. vulgatus, P. copri, B. plebeius, and V. parvula posed essential contributions to the construction of the model. b The ROC curves and Gini values for the UN-risk model constructed at the strain level. B. fragilis ASM181622v2, R. lactatiformans, C. neonatale, F. prausnitzii, and C. aerofaciens ASM987593v1 posed essential contributions to the con-struction of the model

Full size image

Intimated correlations existed between GM, intestinal SCFAs, and nutritional indicators

Based on Mentel analysis, we explored the general impacts of GM, weight to age, and short-chain fatty acids (SCFAs) on nutritional indicators in children, including leukocytes, lymphocytes, vitamin D, Ca, Fe, Zn, etc. (Fig. 5a). We first discovered that the GM was positively correlated with the age (Rpearson = 0.120, Padj = 0.027) and hemoglobin (Rpearson = 0.166, Padj = 0.001). Second, SCFAs were positively associated with the levels of Fe (Rpearson = 0.260, Padj = 0.014) and Zn (Rpearson = 0.155, Padj = 0.042). Third, weight to age, which is an important indicator for malnutrition clinical diagnosis, was positively correlated with other physical indicators, including the white cell number (Rpearson = 0.232, Padj = 0.001), the hemoglobin level (Rpearson = 0.502, Padj = 0.001), the lymphocyte number (Rpearson = 0.266, Padj = 0.001), the serum total protein (TP) level (Rpearson = 0.243, Padj = 0.001), the albumin level (Rpearson = 0.235, Padj = 0.001) and the Ca level (Rpearson = 0.186, Padj = 0.001, Fig. 5a). These findings implied the essential impact of GM and its derived SCFAs on nutrient absorption in humans and disclosed the intimate relationships between nutritional status and various physical indicators.

Fig. 5
figure 5

Correlations among the GM strains, fecal SCFAs, and the UN clinical indicators. a Contributions of SCFAs, Weight to age, and GM on the physiological indicators of UN. The thicker lines represent the higher contributions of variables on the physiological indicators. The oranges, green, and grey lines represent the adjusted P-value < 0.01, 0.01-0.05, and \(\ge 0.05\), respectively. The blue and red boxes represent the positive and negative correlations among the clinical indicators, respectively. b Corre-lation heatmap between the SCFAs and GM strains. c Correlation heatmap between the SCFAs and the physiological indicators of UN. d Correlation heatmap between GM strains and the physiological indicators of UN. The blue and red boxes represent the positive and negative correlations, respectively. *, Padj < 0.05; **, Padj < 0.01; ***, Padj < 0.001

Full size image

Then, we further investigated the detailed associations between GM, SCFAs, physical indicators, and trace elements in the children (Fig. 5b-d). We found that the levels of SCFAs were closely related to GM, e.g., propionic acid was positively correlated with Ruthenibacterium lactatiformans (Rspearman = 0.244, Padj = 0.024), Flavonifractor plautii (Rspearman = 0.222, Padj = 0.041), Eggerthella lenta (Rspearman = 0.230, Padj = 0.034), and Faecalibacterium prausnitzii (Rspearman = 0.273, Padj = 0.012, Fig. 5b). In addition, the associations between different SCFAs and the physical indicators varied in the children (Fig. 5c). Isovaleric acid was significantly positively correlated with the levels of hemoglobin (Rspearman = 0.291, Padj = 0.007) and Fe (Rspearman = 0.238, Padj = 0.028); Valeric acid was positively correlated with the levels of albumin (Rspearman = 0.230, Padj = 0.034) and TP (Rspearman = 0.237, Padj = 0.029); Isobutyric acid was negatively correlated with the numbers of lymphocyte (Rspearman = -0.261, Padj = 0.016) and white cell (Rspearman = -0.229, Padj = 0.035, Fig. 5c). Furthermore, GM is also closely related to physical indicators (Fig. 5d). Among them, Veillonella parvula was positively correlated with vitamin D (Rspearman = 0.230, Padj = 0.034); Clostridium neonatale was negatively correlated with the levels of Zn (Rspearman = -0.214, Padj = 0.050) and Ca (Rspearman = -0.280, Padj = 0.009); and Faecalibacterium prausnitzii was negatively correlated with albumin (Rspearman = -0.341, Padj = 0.001), TP (Rspearman = -0.324, Padj = 0.003), hemoglobin (Rspearman = -0.284, Padj = 0.010), lymphocyte (Rspearman = -0.305, Padj = 0.005), and white cell (Rspearman = -0.267, Padj = 0.014, Fig. 5d). The discoveries provided us with clinical references for the intervention of malnutrition through modulating GM or SCFAs modulations.

Discussion

Numerous studies have investigated the gut microbiome in undernourished children to gain insight into the potential links between gut microbiota and undernutrition. However, these studies often overlook the strain-level diversity within bacterial species, which could provide additional information on the functional roles of specific bacterial strains in the pathogenesis of undernutrition [24, 25]. Our study addresses this gap in knowledge by analyzing the strain-level diversity of the gut microbiome in undernourished children and identifying specific strains that are differentially enriched between undernourished and healthy control children.

Our strain-level analysis revealed that undernourished children had enriched Prevotella copri strains, which have been linked to adverse conditions such as chronic inflammation [26], fat accumulation [27], insulin resistance [28], and glucose intolerance [28]. However, conflicting reports also associate P. copri with improved glucose and reduced visceral fat [29]. Our findings showed that the strains in undernourished and healthy children were from different clades, potentially having different metabolic roles. Higher taxonomic levels can obscure candidate bacteria that contribute to human diseases, and enriched pathogens at the strain level provide new clues to the pathogenesis of undernutrition.

We also highlighted the identified differences in GM composition between the UN and HC group, particularly the absence of certain beneficial bacterial strains in the UN group, highlighting potential associations between specific bacteria and undernutrition status. For instance, Bacteroides fragilis and Bacteroides uniformis strains are known to be involved in the fermentation of complex polysaccharides and the production of SCFAs [30,31,32]. Polysaccharide and oligosaccharide metabolism provides nutrition and vitamins to the host and other intestinal microbial residents [33]. SCFAs, such as butyrate, acetate, and propionate, serve as important energy sources for colonocytes, maintain intestinal homeostasis, and contribute significantly to host nutrition [33, 34]. Otherwise, Faecalibacterium prausnitzii is also one of the most important butyrate-producing bacteria [30]. The absence of B. fragilis, B. uniformis, and F. prausnitzii in undernourished children might result in inadequate production of SCFAs and reduced efficiency in polysaccharide metabolism. B. fragilis ASM181622v2 might be one of the key strains. Our results also showed that F. prausnitzii was positively correlated with propionic and negatively correlated with albumin, TP, hemoglobin, lymphocyte, and white cells. This bacterium has potential roles in maintaining gut health. Moreover, the presence of lower levels of Bifidobacterium pseudocatenulatum strains, which are well-known for their beneficial effects on host health, in the UN group might suggest that undernourished children experienced impaired fermentation of dietary fibers, reduced production of short-chain fatty acids, and compromised immune function [30, 35,36,37]. These findings have significant implications for addressing the increasing rise of malnutrition. Understanding the specific roles of these bacteria in nutrient metabolism and gut health can potentially lead to interventions targeting the gut microbiome to alleviate undernutrition. Strategies such as dietary modifications (e.g., increasing dietary fiber and complex carbohydrates), probiotic supplementation with specific beneficial bacteria, or fecal microbiota transplantation may be explored to rebalance the gut microbiota and improve nutritional outcomes.

Co-occurrence network analysis revealed that different topological characteristics between UN and HC groups might be linked to different inter-species interactions or system efficiency [38]. Although the gut microbiota networks in both groups had scale-free properties, the hub nodes, which play a critical role in preserving the overall functionality of the network, were different [39,40,41]. Notably, the gut microbiota network of undernourished children exhibited higher clustering coefficient and modularity, along with lower graph density, lower average degree, and longer average path length. These characteristics are associated with lower community stability, worse microbial cooperation, and decreased communication [38, 40, 42].

Moreover, we constructed a UN risk model based on the strain compositions, demonstrating a better predictive effect than the model based on species compositions. At the species level, Bacteroides vulgatus, Bacteroides uniformis, and Prevotella copri were critical contributors to the risk prediction model and differentially enriched between undernourished and healthy children. However, at the strain level, their high intraspecific diversity resulted in no differentially enriched strains between the two groups. Critical strain biomarkers with significantly different abundances were Bacteroides fragilis ASM181622v2, Ruthenibacterium lactatiformans, and Clostridium neonatale sp., with critical biological functions in intestinal infection or homeostasis maintenance [43]. Our strain-level analysis provided a more precise UN risk model, enhancing our understanding of the associations between childhood undernutrition and GM.

Conclusions

In conclusion, our study explored the characteristics of GM and its co-occurrence networks in undernourished children from the strain level perspective, supplementing the understanding of the GM’s roles in the pathogenesis of undernutrition, providing a more precise UN risk model, and laying the groundwork for undernutrition interventions.