Non-stochastic reassembly of a metabolically cohesive gut consortium shaped by N-acetyl-lactosamine-enriched fibers

ABSTRACT

Diet is one of the main factors shaping the human microbiome, yet our understanding of how specific dietary components influence microbial consortia assembly and subsequent stability in response to press disturbances – such as increasing resource availability (feeding rate) – is still incomplete. This study explores the reproducible re-assembly, metabolic interplay, and compositional stability within microbial consortia derived from pooled stool samples of three healthy infants. Using a single-step packed-bed reactor (PBR) system, we assessed the reassembly and metabolic output of consortia exposed to lactose, glucose, galacto-oligosaccharides (GOS), and humanized GOS (hGOS). Our findings reveal that complex carbohydrates, especially those containing low inclusion (~1.25 gL⁻¹) components present in human milk, such as N-acetyl-lactosamine (LacNAc), promote taxonomic, and metabolic stability under varying feeding rates, as shown by diversity metrics and network analysis. Targeted metabolomics highlighted distinct metabolic responses to different carbohydrates: GOS was linked to increased lactate, lactose to propionate, sucrose to butyrate, and CO₂, and the introduction of bile salts with GOS or hGOS resulted in butyrate reduction and increased hydrogen production. This study validates the use of single-step PBRs for reliably studying microbial consortium stability and functionality in response to nutritional press disturbances, offering insights into the dietary modulation of microbial consortia and their ecological dynamics.

Introduction

The assembly of the infant gut microbiome can have a dramatic lifelong impact. Crucial elements like delivery mode, vertical transmission from caregiver to baby, feeding, and other environmental cues like family members and pets have been shown to alter the microbiome.^1,2 For example, breast milk not only primes infants for the introduction of solid foods after weaning but also imparts a lasting impact, potentially explaining the high levels of memory T cells observed in breastfed infants through adolescence.^3,4 This immunological resilience translates into greater resistance against infections.^5,6 Conversely, disruptions during early gut colonization, particularly in the context of C-section versus vaginally delivered infants, can induce imbalances with repercussions for immune function, predisposing individuals to allergies and inflammatory conditions.^7–9 Moreover, once the infant gut microbiome undergoes its critical assembly phase at approximately 3 years of age,¹⁰ making effective and stable modifications becomes difficult.

In vitro models are valuable for studying the responses of the gut microbiome under controlled conditions, especially given the challenges associated with studying dynamic microbial communities in vivo. Few studies employing continuous flow bioreactor systems have concurrently evaluated microbial taxonomic responses and associated metabolic outputs resulting from dietary changes, highlighting a significant gap in our understanding.^11,12 In the host, as most soluble monosaccharides are absorbed in the small intestine, the colonic microbiota predominantly engages in fermentative activity due to reduced oxygen availability. There, colonic microbiota’s ability to extract energy from highly reduced carbohydrates is restricted under those conditions. Dietary carbohydrates with specific structural complexities, such as prebiotic fibers and human milk oligosaccharides (HMOs) found in breast milk, reach the lower intestine, and undergo selective fermentation into short-chain fatty acids (SCFAs).¹³ SCFAs are shared among consortium members through cross-feeding, a phenomenon that may play a critical role in determining the overall stability and resilience of the consortium. Linking taxonomic responses and metabolic outputs provides a nuanced understanding of the intricate dynamics within microbial consortia, thereby contributing to a more holistic comprehension of microbial ecology in the context of disruptive events, such as pulse disturbances (abrupt, short-term events) or press disturbances (persistent, prolonged events).

Priority effects, which refer to the influence of early colonizing species on the subsequent assembly and stability of microbial communities, can play a significant role in determining the long-term composition and resilience of gut microbiota.¹⁴ Early-arriving species may either facilitate or inhibit the establishment of later-arriving taxa, potentially shaping the overall community structure.¹⁵ However, in this study, priority effects were not explored due to the use of a standardized inoculum, which provided a uniform starting point across experimental conditions. Furthermore, our analysis focused exclusively on steady-state taxonomic and metabolic data, omitting the transient assembly process that typically reveals priority effects.

Continuous flow bioreactor systems offer the ability to conduct multiple experimental replicates under highly controlled and standardized conditions.^16–18 By simulating key natural environmental conditions in the gut, bioreactors can facilitate consortium-level research on dynamic microbe–microbe interactions. One of the first such dynamic systems was the Simulated Human Intestinal Microbial Ecosystem (SHIME) reactor, developed by Willy Verstraete’s group at the University of Ghent in 1992.¹⁹ The authors generated fair correlations with in vivo data and later provided additional validating data, demonstrating the SHIME reactor’s suitability as a simulation when inoculated with fecal material from eight healthy adults.¹⁹ Other research groups have tested similar multi-stage dynamic systems, consistently achieving comparable results by using a fixed culture medium supply rate.^20–24 Single-step packed-bed reactors (PBRs) are a simpler type of system than those discussed above with ideal characteristics for reproducible simulations to evaluate consortium assemblies, making them ideal for simultaneously evaluating metabolomic fingerprints.²⁵ To date, few studies have successfully linked steady-state complex community-level taxonomic data with metabolomics in response to changes in dietary composition and availability.

Only one publication has tested PBRs for culturing human gut microbial consortia.²⁶ The study employed a twin-vessel single-stage chemostat model of the human distal gut with a simulated mucosal environment and aimed to determine the stability of planktonic and biofilm consortia over time and their responses to antibiotic perturbation. The study revealed an enrichment of Firmicutes (now Bacillota) and a decreased relative abundance of Bacteroidetes (now Bacteroidota) in biofilm samples compared to planktonic consortia. Notably, this culturing method successfully retained Akkermansia muciniphila, a bacterium that utilizes mucus and is commonly present in healthy mucosa, validating the model. The authors demonstrated the reproducibility of simulated mucosal biofilms, their stability over time, and the differential impact of clindamycin-induced perturbation on biofilm consortia formed during and after exposure. This study provides a relevant framework for employing a dynamic, continuous in vitro system for studying the effects of pulse disturbances, in this case, antibiotic impact on a fully assembled consortium derived from a singular healthy adolescent donor.

In this study, we employed a single-step PBR to assess the impact of two sources of press disturbances in re-assembling a microbial consortium. The same pooled infant stool samples were independently inoculated and cultured in media, only differing in the carbohydrate composition and complexity. Two independent reactors for each condition were tested, including carbohydrate compositions containing a compound found in human breast milk (LacNAc). The second experimental source of press disturbance applied to the above reactors was the progressive increase in feeding rates. Particularly, we focused on the presence of LacNAc, one of the most abundant building blocks in HMOs. We have previously shown using a mice model that these compounds can selectively promote the growth of Bifidobacterium species, which are associated with health-promoting effects.²⁷ While these structural building blocks are recognized for their structural complexity and the specialized enzymatic capacity required for microbial utilization, it remains unclear their impact on the broader community-level cross-feeding relationships and, consequently, on community resilience and cohesiveness.

The main objective of our study was to assess the impact of persistent dietary disturbances (press disturbances) on the reassembly of infant-derived microbial consortia. We iteratively explored how dietary compounds – such as sucrose, lactose, galacto-oligosaccharides (GOS), and LacNAc-enriched GOS determine microbial community structure and stability on a fully assembled consortium. Finally, to further mimic the selective pressure within the gastrointestinal tract, we incorporated bile salts into the media. We hypothesized that carbohydrate compositions containing complex, low-inclusion compounds like LacNAc would enhance the stability, metabolic cohesiveness, and resilience of the reassembled microbial consortium.

We postulated that more structurally complex carbohydrates, such as LacNAc, would lead to reduced microbial diversity while increasing cohesiveness and stability in the assembled community. In this context, “cohesiveness” pertains to the degree of stability in the relative abundance of OTUs within a consortium. Specifically, we evaluated how consistently these taxonomic units maintained their abundance in response to disturbances, indicating a balanced and stable community where fluctuations in population dynamics are minimized. Additionally, we expected that faster feeding rates would further impact community stability and metabolic outputs, indicating resilience or lack of in response to resource availability.

Materials and methods

Inoculum preparation

To avoid variability due to inconsistent inoculum source, our experiments employed a standardized, pooled inoculum derived from fecal material from three healthy one-year-old infants in a clinical setting (University of North Carolina at Chapel Hill IRB# 12–0809) whose characteristics have been described previously.²⁸ Fecal samples from the infants were collected as previously described.²⁹ Using a pooled inoculum served the dual purpose of capturing maximum microbial diversity and ensuring adequate cell viability. Briefly, parents collected approximately 200 mg of fecal material from a single diaper and immediately placed it in a tube completely submerged in 10 mL of thioglycolate broth (BD Difco™ Fluid Thioglycolate Medium Catalog # 225650), a multipurpose, enrichment medium that produces a range of oxygen concentrations along its depth and returned it through overnight shipping. Once received, samples were stored at −80°C until analysis. To avoid inoculum freezing and thawing, we created a working stock containing the highest initial microbial diversity. We first combined all fecal samples under anaerobic conditions and divided them into 0.5 mL aliquots for storage at −80°C. To preserve the cell viability of anaerobes, all processes were performed in a Coy chamber (Coy Laboratory Products, Inc, Michigan, USA) under an anaerobic gas mix (5% CO₂, 10% H₂, 85% N₂).

Culture medium

The basal medium used in this study contained 10 g L⁻¹ yeast extract (BD Bacto, Lot# 6088514; Becton, Dickenson & Co.), 0.05 g L⁻¹ MgSO₄·7 h₂O, 2.5 g L⁻¹ (NH₄)₂HPO₄, and 0.005 g L⁻¹ MnSO₄·H₂O. Dietary carbohydrate compositions were always added at a final concentration of 20 g L⁻¹. The evaluated carbohydrate compositions were lactose, glucose, and galacto-oligosaccharides (GOS) (Oligomate, Yakult Pharmaceutical Industry Co., Ltd.). Two additional experiments were performed with GOS (Oligomate) and hGOS (Oligomate containing 1.2 g L⁻¹ LacNAc) where 0.4 g L⁻¹ of bile salts #3 (45–55% sodium cholate, 45–55% sodium deoxycholate, Hardy Diagnostics) were added to the culture medium (bile salt composition confirmed via NMR analysis, METRIC at NC State University, data not shown). All bioreactor experiments were carried out in duplicate.

Packed-bed reactor setup

The bioreactors consisted of custom-made water-jacketed glass reactor vessels (6.5 cm internal diameter × 23.5 cm height, 600 mL volume; manufactured by Quark Glass, New Jersey, USA) connected to Biostat B-Plus control units (Sartorius Stedim Biotech S.A., Germany). The reactor vessels were packed with porous glass-foam beads 4–8 mm diameter (Dennert Poraver GmbH, Schlüsselfeld, Germany) to provide the system with a high surface area and low metabolic gas retention. A recirculation loop was added to the system to facilitate homogeneous mixing, which increased the total volume to 620 mL. This recirculation loop contained a sampling port, a connector through which fresh media was fed, and a connector through which a 10% NaOH solution was provided to control pH (5.8). The porosity of the packed bed was 40%, and the active volume of the entire system was 350 mL.²⁸ The ascending velocity of the recirculation loop in the reactor was 281.7 cm h⁻¹, indicating a very well-mixed system.²⁸ Before inoculation into bioreactors, the medium was depleted of oxygen to allow for the proliferation of anaerobic and slow-growing organisms. Fifteen mL of liquid were sampled from a port in the recirculation loop of the reactor. Samples were mixed thoroughly, and 1.5 mL aliquots were placed into 2 mL microcentrifuge tubes and centrifuged at 10,000 × g for 2 min. The supernatants were transferred to a 2 mL vial for HPLC analysis, and pellets were stored at −80°C for microbiome analysis.

Packed-bed reactor sampling scheme

The PBRs were inoculated with pooled fecal samples as described above and cultured under conditions reflecting the physiological ranges of the intestinal environment (Table 1). We used a dynamic feeding strategy that involved stepwise decreases in retention times (RT) and minimal media containing either sucrose, lactose, or galacto-oligosaccharides (GOS). Additional experiments were conducted feeding GOS supplemented with bile salts (GOS + BS) and GOS enriched with LacNAc, hereafter referred to as humanized GOS [hGOS] also in the presence of bile salts (hGOS + BS). This resulted in five total carbohydrate source conditions performed with two independent experiments for a total of ten. Each condition replicated the same scheme of continuous operation testing four incremental feeding rates, e.g. step decreases of retention times (RT). The feeding rates, represented by the stepwise increments in Figure 1, were chosen to mimic four estimated colonic transit times for infants, which ranged from 6 to 25 h.²⁹ Following this sampling scheme, the initial inoculum is cultured through 20 volume replacements in each experiment. This ensures that any potentially spurious sequences are significantly diluted and eventually eliminated from the system over time ([Y]_t = $\left(\frac{C_{0\ast V_0}}{V_R}\right)e^{-\mathrm{Dt}})$, as each volume replacement further reduces their presence, a fundamental property of dynamic continuous systems.³⁰

Table 1.

Packed-bed reactor operational parameters.

Media Feeding (Dilution) Rate (h−1)	Flow Rate (mL h−1)	Retention Time (h)	Carbohydrate Component Feeding Rate (g h−1)
0.04	14.0	25.0	0.28
0.08	28.0	12.5	0.56
0.12	42.0	8.30	0.84
0.16	56.0	6.25	1.12

Figure 1.

Reactor setup (left), experimental design, and sampling strategy (right). Five dietary carbohydrate compositions were tested, each evaluated in duplicate using single-step packed-bed reactors (PBRs). Each PBR was inoculated with the pooled fecal sample from healthy one-year-old infants serving as the inoculum and maintained under constant conditions. Within each independent experiment replicate, four feeding rates were tested incrementally to simulate different retention times (RT) represented by the staircase shape. At each feeding rate, biological samples were collected at steady-state over one retention time, resulting in three samples per feeding rate. Ten independent PBR experiments were conducted to assess the impact of the dietary compositions on the stable assembly of an infant gut microbial consortium. For each of the 120 samples generated, microbial diversity and metabolite concentrations of the assembled consortia were determined through 16S rRNA gene amplicon sequencing and targeted HPLC analysis (see materials and methods for details).

Each of the PBRs loaded with medium containing the dietary carbohydrate composition to be evaluated was consistently inoculated and cultured in batch mode for 48 h before initiating feeding. Once feeding commenced at an initial rate of 0.04 h⁻¹, it was consistently determined that the system required a total of 192 h, the equivalent of 8 volume replacements, to achieve a steady-state condition independently of dietary carbohydrate compositions tested. During this transient period, the trajectory of the microbial consortium composition underwent an adaptation period to the dietary compositions being supplied (data not shown). After 192 h, each steady state condition was corroborated by monitoring in-line the real-time generation of CO₂ and H₂. Additionally, we determined the concentrations of specific short-chain fatty acids (SCFAs) offline, using also these concentration values as indicators of steady-state conditions and indicators of a balanced consortium. Biological samples were collected at intervals of at least one retention time for a total of three samples per feeding rate before adjusting the feeding rate (Figure 1). This resulted in a total of 120 samples (5 compositions × 2 independent experiments × 4 feeding rates × 3 biological samples per feeding rate = 120 samples).

Targeted metabolite and gas analysis

CO₂ and H₂ generated by the consortium were continuously monitored in-line with an ABB (EL3020; Zürich, Switzerland) gas analyzer and a Pfeiffer Omnistar (Pfeiffer Vacuum, USA) mass spectrometer connected in series. Nitrogen was supplied into the system at a rate of 0.208 L min⁻¹ and used to maintain anaerobiosis and as a stripping carrier gas. The system was determined to be at a steady state when the values of the in-line gas measurements remained constant for at least three retention times. Steady-state was corroborated by the compositional stability of short-chain fatty acids (SCFAs: acetate, propionate, and butyrate) and lactate. Measurements of targeted metabolites were performed by HPLC (Agilent 1100 chromatograph) with a refractive index detector under isocratic conditions. Metabolite separation was performed using 5 mm H₂SO₄ at 0.5 mL min⁻¹ with a Phenomenex Rezex ROA column at 65°C.

DNA isolation and 16S rRNA gene amplicon sequencing

DNA Isolation was performed as described previously.^26,31 Briefly, samples were transferred to a 2 mL tube containing 200 mg of 106/500 μm glass beads (Sigma, St. Louis, MO) and 0.6 ml of Qiagen ATL buffer (Valencia, CA), supplemented with 60 mg mL⁻¹ lysozyme (Thermo Fisher Scientific, Grand Island, NY). The suspensions were incubated at 37°C for 1 h with occasional agitation. Subsequently, the suspension was supplemented with 600 IU of Qiagen proteinase K and 0.3 mL of Qiagen AL buffer followed by a 70°C incubation for 1 h. After a brief centrifugation, supernatants were aspirated and transferred to a new tube containing 0.5 mL of ethanol. DNA was purified using a standard on-column purification method with Qiagen buffers AW1 and AW2 as washing agents and eluted in DNase free water.^32–34 A total of 12.5 ng of DNA was used for 16S rRNA gene amplicon sequencing. Total DNA was amplified using universal primers targeting the V4 region of the bacterial 16S rRNA gene as described.^32,35 Primer sequences contained overhang adapters appended to the 5’ end of each primer for compatibility with the Illumina sequencing platform. The primers used were F515/R806.^36,37 Each 16S rRNA gene amplicon was purified using the AMPure XP reagent (Beckman Coulter, Indianapolis, IN). In the next step, each sample was amplified using a limited cycle PCR program, adding Illumina sequencing adapters and dual‐index barcodes (index 1(i7) and index 2(i5)) (Illumina, San Diego, CA) to the amplicon target. The thermal profile for the amplification of each sample had an initial denaturing step at 95°C for 3 min, followed by a denaturing cycle of 95°C for 30 s, annealing at 55°C for 30 s and a 30-s extension at 72°C (8 cycles), a 5-min extension at 72°C and a final hold at 4°C. The final libraries were again purified using the AMPure XP reagent (Beckman Coulter), quantified, and normalized before pooling. The DNA library pool was then denatured with NaOH, diluted with hybridization buffer, and heat-denatured before loading on the MiSeq reagent cartridge (Illumina) and the MiSeq instrument (Illumina). Automated cluster generation and paired-end sequencing with dual reads were performed according to the manufacturer’s instructions.

Sequence analysis, including Chimera removal, Operational Taxonomic Unit (OTU) clustering, taxonomy assignments, and alpha diversity calculations using the Shannon Index (SI), was performed in QIIME2 (v2019.10; https://qiime2.org). Sequence reads were analyzed without using the DADA2 algorithm to filter sequences.^38,39 Instead, sequences were demultiplexed, joined, and dereplicated using the q2 vsearch plugin⁴⁰ and subsequently filtered with the q2 quality-filter q-score algorithm using default parameters.⁴¹ After quality filtering, denoising, and chimera removal, a total of 11,286,641 sequences were retained across all samples, with a mean sequencing depth of 66,784 reads per sample (range: 2,726–172,684). OTUs were then clustered de novo based on a genetic similarity threshold of 97% before assigning taxonomy using the Greengenes Database⁴² (version 13.8) as a reference in QIIME2. OTUs were filtered based on the number of reads rather than relative abundances, with a minimum threshold of 200 reads required for an OTU to be included in downstream analyses.

Construction and annotation of phylogenetic trees

Before generating phylogenetic trees, the representative sequences generated in QIIME2 were condensed using the condense_workflow.py script as implemented in the PhyloToAST (v1.3.0) package.⁴³ The condensed set of representative sequences was aligned using the Multiple Alignment using Fast Fourier Transform (MAFFT) program, and a tree was generated from this alignment using the RaxML algorithm.^44,45 Finally, the final tree was annotated with relative abundance data using the iTol.py script in PhyloToAST and visualized using the Interactive Tree of Life browser tool.^43,46

Co-occurrence network analysis

Networks were generated using the SCNIC (Sparce Co-occurrence Network Investigation for Compositional Data) plugin for QIIME2 (https://github.com/shafferm/q2-SCNIC). Data were filtered using the built-in method, which removes all samples with a total OTU relative abundance of less than 500 and any OTUs with fewer than two occurrences. OTU correlation tables were calculated for each dietary carbohydrate composition using the Kendall-Tau correlation method in SCNIC. The correlation network was then constructed based on a p-value cutoff of p < 0.05. The resulting network was plotted and annotated using Cytoscape (www.cytoscape.org).

Statistical analysis

All statistics were performed in OriginPro 2015 (OriginLab Corporation, Northampton, MA). Data were first assessed for normality and homogeneity of variance using the methods of Shapiro-Wilk and Levene, respectively. Normally distributed data were subject to an ANOVA followed by Tukey’s HSD to compare means unless otherwise noted. Alpha diversity was calculated using the default QIIME2 diversity command which applies the core-metrics phylogenetic method that rarifies an OTU table to a maximum depth of 5,000 reads/sample before computing alpha diversity metrics. Unweighted DPCoA plots and the analysis thereof were performed in R 3.6 using the phyloseq, ggplot2, and vegan packages using Bray-Curtis dissimilarity-index methods.^47–50 Beta diversity statistical analyses were performed using the PERMANOVA, and PERMDISP functions of the vegan package in R. Differential abundance plots were generated using the DESeq2 package⁵¹ where the OTU table was pruned to include only taxa for which the sum of counts in each sample is greater than 0. Unless otherwise noted, the alpha for all statistical tests was set at 0.05. The metabolic outputs for each carbohydrate source condition were analyzed using pairwise statistical tests, specifically, the Student’s t-test for normally distributed data and the Mann–Whitney U test for non-normally distributed data, as provided by the M2IA bioinformatics platform.⁵²

Carbohydrate complexity shapes the composition of microbial consortia assembly

We tested the impact of five carbohydrate compositions on community reassembling using a pooled stool inoculum. The phylogenetic analysis was conducted using 16S rRNA gene sequences of 120 samples from in-vitro culturing (Table S1). Figure 2 shows the presence/absence of unique OTUs identified in each carbohydrate source across feeding rates, including those initially detected in the inoculum. Of these OTUs, 118 out of 219 (53.9%) were detected in the inoculum, with diminishing numbers in reactors fed with GOS (106/219, 48.4%), sucrose (100/219, 45.7%), lactose (88/219, 40.2%), hGOS + BS (76/219, 34.7%), and GOS + BS (69/219, 31.5%).

Figure 2.

Overall phylogenetic tree displaying experimental OTUs present or absent. The internal colored circle shows the phylogenetic tree of representative sequences found at least in one of the 125 samples evaluated. The phylogenetic tree was generated in QIIME using the RaxML algorithm and annotated with QIIME2-sourced presence-absence data. PhyloToAST and iToL were utilized to represent external-colored dots indicating the presence or absence of each representative sequence for a given OTU when cultured in each dietary carbohydrate source and condition. The grayscale strip just outside the tree leaves represents OTU persistence, which refers to the presence or absence of specific OTUs across all experimental conditions. Black indicates that an OTU was present in 100% of the experimental conditions, while the lightest shade represents OTUs found in only a small fraction of the conditions.

Thirty-nine unique OTUs were detected in the inoculum only, including taxa associated with health-related properties such as Bacteroides caccae, Clostridium butyricum, Roseburia faecis, and Blautia obeum .^53–55 Conversely, Lacticaseibacillus zeae, Streptococcus alactolyticus, and Clostridium septicum were not detected in the inoculum despite being found in every condition except hGOS + BS. The specific anaerobic requirements of these taxa suggest that they may not been in sufficient numbers in the inoculum.

The distribution of specific taxa across culturing conditions revealed distinct patterns. For example, Bifidobacterium spp. was detected consistently in all, while B. bifidum was present exclusively in sucrose, lactose, and GOS, with a notable absence in lactose and GOS at the highest feeding rate tested (D = 0.16 h⁻¹). Conversely, B. adolescentis was found in all conditions except GOS, lactose, and sucrose. These distinct patterns highlight species-specific responses to varying carbohydrates and availability (feeding rates), contributing to a nuanced understanding of the ecological dynamics within the experimental framework.

In the experimental condition of hGOS + BS, representative of basal media supplemented with GOS, LacNAc, and bile salts, we discerned a distinctive taxonomic profile not observed in alternative experimental conditions. Notably, Porphyromonas endodontalis, Prevotella (formerly Bacteroides) intermedia, Leptotrichia spp., Eikenella spp., and Aggregatibacter spp. were exclusively identified within this specific condition at a feeding rate of 0.12 h⁻¹. Indeed, this feeding rate in the hGOS + BS condition had the highest number of OTUs detected of all feeding rates (47/219, Table S1).

Figure 2 highlights other keystone species (in red), known from the literature to play crucial roles in gastrointestinal health. These species were present in at least 75% of conditions, with exceptions such as Akkermansia muciniphila which was initially present in the inoculum but became exclusive to hGOS + BS at a feeding rate of 0.12 h⁻¹. Other exceptions include Alistipes putredinis and Ruminococcus bromii which were both exclusively found in the inoculum whereas Faecalibacterium prausnitzii and Thomasclavelia ramosa were consistently found in all carbohydrate sources and feeding rates.

Carbohydrate feeding rate impacts climax consortium assembly and composition

Diversity metrics, including the number of overall observed OTUs and Shannon Index (SI) across all experiments, concur within the expected range of values previously reported for one-year-old infants²⁷ (Figures 3 and S1). The inoculum had the highest diversity among the tested substrates and feeding rates, except for sucrose and GOS at a rate of 0.16 h⁻¹. Likewise, the 0.04 h⁻¹ feeding rate, which corresponds to a retention time of 25 h, consistently showed the lowest diversity across all substrates (Figures 3a-e), with only sucrose and GOS conditions showing increasing diversity as feeding rates increased. The addition of bile salts to GOS and hGOS resulted in a marked diversity reduction.

Figure 3.

Alpha diversity metrics (Shannon Index) were calculated in QIIME2 from steady-state samples at each feeding rate tested and across substrates. Panels A-E display the values obtained for each dietary carbohydrate structure and condition (sucrose, lactose, GOS, GOS + BS, and hGOS + BS experiments, respectively). Panels F-I compare the SI values obtained for each dietary carbohydrate structure for each feeding rate tested (D= 0.04, 0.08, 0.12, 0.16 h⁻¹, respectively). Each feeding rate was tested in duplicate experiments, with triplicate samples collected per experiment, resulting in 6 samples per feeding rate. Each metric was iteratively calculated 10 times for each sample for 60 data points represented in each boxplot, except for the inoculum which contains 20 data points from sequencing in duplicate. This iteration accounts for variability and ensures the robustness of the diversity metric. The mean for each boxplot is represented by a small square, while a line represents the median. Pairwise comparisons that share a letter above the boxplot are considered non-significant. Those not sharing a letter are considered significantly different (p < 0.05).

A significant decrease in the number of OTUs was also observed when comparing the inoculum and the subsequent feeding rates for all conditions, except for GOS+BS (Figure S1). The number of OTUs remained stable at subsequent feeding rates suggesting that the differences in diversity between feeding rates were driven by the communities’ distribution of species, rather than their number of different taxa.

Beta diversity metrics provided insights into the relationships between reassembled microbial consortia from each dietary carbohydrate structure, each steady-state, and the selective pressure from bile acids. A Principal Coordinates Analysis (PCoA) plot (Figure 4a and Table 2) showed a tight community cluster of samples from the lactose group but overlapped GOS and sucrose clusters, probably due to the composition of the GOS preparation used in the study, which includes free sugars (lactose, galactose, and glucose) that can influence microbiome composition. Likewise, the selective impact of bile salts on the culture medium containing either GOS (GOS + BS) or GOS enriched with LacNAc (hGOS + BS) was observed as these conditions clustered together and away from sucrose, lactose, and GOS. The overall PERMANOVA and all pairwise comparisons indicated significant differences across the five experimental conditions (p < 0.005).

Figure 4.

Diversity and differential abundance analyses were calculated using the phyloseq, vegan, and DESeq2 packages in R, respectively. Panel a represents the unweighted PCoA ordination of OTU relative abundance data from each sample taken at a steady state at each feeding rate. Colored ellipses represent 95% confidence intervals for all data in each dietary carbohydrate composition experiment. Feeding rates (dilution rates) and carbohydrate composition (experiments) are represented by unique shapes and colors, respectively. Panel B represents differential abundance analyses that were performed on QIIME2-sourced OTU tables. Positive log₂ values indicate a higher relative abundance in the reactor compared to the inoculum, while negative values indicate a lower relative abundance in the reactor compared to the inoculum. Only OTUs that displayed statistically significant differences (p < 0.05) were shown by decreasing significance from left to right. The presence of multiple data points reflects variations in relative abundances, as some OTUs were originally identified at the species level, while others were only resolved at the genus level.

Table 2.

P-values from PERMANOVA and PERMDISP tests performed comparing all feeding rates for each given carbohydrate structure using the vegan package in R.

Carbohydrate Source	PERMANOVA (By feeding rate)	PERMDISP (By feeding rate)	F-Value
Lactose	0.001	0.001	33.763
Sucrose	0.264	0.001	19.4
GOS	0.001	0.001	14.771
GOS + BS	0.012	0.316	5.0601
hGOS + BS	0.838	0.003	4.393

A significant PERMANOVA indicates that the spatial median among groups is different. In contrast, a significant PERMDISP indicates that the within-group dispersion among groups is different.

PERMANOVA analysis across feeding rates within each tested substrate showed significant clustering when lactose, GOS, and GOS+BS were used as substrates (F > 5, p < 0.05) (Table 2), but not hGOS or sucrose. When we carried out a PERMDISP analysis, which characterizes the distance between each sample and the centroid of the group to which it belongs, only GOS+BS had a non-significant value suggesting that the centroid of the group drove the dispersion of the samples.

We next sought to identify taxa that showed a significantly increased or decreased relative abundance compared to the inoculum in each substrate regardless of the feeding rate (Figure 4b). A decrease in the relative abundance of Bacteroides, Clostridium, Alistipes, Roseburia, Dorea, and Blautia was observed across all nutritional conditions compared to the inoculum. Reductions in Faecalibacterium were observed under all conditions except sucrose. Parabacteroides and Ruminococcus were reduced in sucrose, lactose, and GOS but not in the groups where bile salts were included. Likewise, Bacillota, Enterococcus, Bifidobacterium, and Streptococcus were increased in all conditions, except for the conditions where bile salts were added. Only GOS showed decreased Odoribacter and increased Stenotrophomonas and Bacillus, while Weissella showed an increase in both GOS and lactose.

Complex carbohydrate compositions determine dynamic microbial network cohesion and stability

Microbial networks for each carbohydrate, including all the samples obtained at steady states across each feeding rate, were generated using SCNIC⁵⁶ (Figure 5 and Figures S2–5). When a network represents nodes as OTUs or bacterial taxa, as illustrated in Figure 5, the number of nodes and edges, or the association between pairs of OTUs that significantly shift in relative abundance, are often associated with the “keystoneness” of an OTU.⁵⁷ This concept denotes the extent to which a node contributes to the overall cohesiveness and stability of the microbial consortium within the network.⁵⁷ Co-occurrence network analysis generates several other key parameters, including closeness centrality (indicating the proximity of a node to all other nodes in the network) and degree (the number of connections a node has with other nodes). In Figure 5, the node sizes visually represent the closeness centrality while the node color corresponds to the degree of each node.

Figure 5.

Co-occurrence network of dynamic overall live consortium members interactions while testing hGOS + BS carbohydrate compositions and feeding rates. Data were generated in QIIME2 using the q2-scnic plugin. Node size and node color represent closeness centrality and degree, respectively. Nodes are ordered by increasing degrees in a clockwise fashion. Edge width and color represent adjusted p-value (significance) and R-value (correlation). The bar graphs adjacent to each node represent the number of positive outgoing interactions (green bars) or negative (red bars) interactions for each node by feeding rate. Nodes that do not have a corresponding graph did not have any outgoing interactions.

The hGOS+BS network displayed the lowest node count with 22 nodes (Figure 5), in contrast to other conditions, where the node counts were 37, 53, 54, and 61 for GOS+BS, lactose, sucrose, and GOS, respectively (Table 3). Furthermore, the hGOS+BS network exhibited the highest heterogeneity (0.566), representing the variance of the degree distribution (number of interactions per node),⁵⁸ in comparison to the other conditions tested (0.298, 0.310, 0.359, and 0.372 for the lactose, GOS+BS, GOS, and sucrose compositions, respectively; Table 3). High network heterogeneity indicates that within the network, there is a wide variety of ways in which nodes are connected which suggests the network does not follow a uniform pattern of connections. Instead, some nodes such as Thomasclavelia ramosa, Faecalibacterium prausnitzii, and Companilactobacillus paralimentarius acted as connectivity hubs, corroborating their role as keystone species, while others maintained fewer connections (Figure 5). This variation in connectivity highlights the complex interaction landscape within the network and underscores the pivotal roles certain species play in maintaining the structural and functional integrity of the ecosystem.

Table 3.

The number of nodes with significant interactions and the total number of significant interactions across feeding rates.

Carbohydrate Source	Number of Unique OTUs Present Simultaneously	Number of Nodes w/Significant Interactions	Total Number of Significant Interactions	Network Heterogeneity
Lactose	68	53	577	0.298
Sucrose	78	54	790	0.372
GOS	79	61	825	0.359
GOS + BS	46	37	252	0.310
hGOS + BS	47	22	67	0.566

Only nodes that change significantly (p < 0.05) in relative abundance across all four feeding rates are counted. Significant interactions are counted when the change in relative abundance of two different OTUs is significantly correlated.

In our examination of microbial diversity within the hGOS + BS experiment (Figure 2), 79 unique OTUs were identified. However, at any given feeding rate, the presence of OTUs was limited to a maximum of 47 simultaneously. This discrepancy prompted a detailed investigation into the specific feeding rate conditions that might facilitate or hinder the co-occurrence of OTUs, aiming to elucidate the dynamics of positive and negative interactions among them. These are represented by the bar graphs beside each node in Figure 5. Interestingly, Faecalibacterium prausnitzii exhibited high closeness centrality and degree, with many significant edges, yet only showed positive interactions with Ruminococcus gnavus and Clostridium spp. at the lowest feeding rate. Conversely, Bacteroides spp. was identified in all feeding rates but exhibited negative interactions with Companilactobacillus paralimentarius, Proteus spp., and Ruminococcus gnavus also at the lowest feeding rate. These observations suggest specialized relationships favorable only under certain resource availability.

Metabolic fingerprints of short-chain fatty acids from each climax consortium reveal non-stochastic responses to dietary carbohydrates

We conducted targeted metabolomics to determine changes in the generation of lactate, SCFAs, and gases in response to carbohydrates and feeding rates (Figure 6). Sample analysis confirmed the complete utilization of the dietary carbohydrate compositions for each feeding rate tested, with overall stoichiometric carbon recovery approaching 100% (data not shown). The inter-experimental variance of the targeted metabolites’ concentrations decreased as follows: lactose → sucrose → GOS → GOS+BS → hGOS+BS. A PCoA plot revealed that sucrose feeding was significantly associated with increased butyrate and CO₂ production, GOS was associated with lactate, lactose was associated with propionate, and the presence of bile salts with increased H₂ generation (Figure 6).

Figure 6.

Results of the PCA for all experiments and their respective metabolomic data. Ellipses represent 95% confidence intervals for centroid position. Feeding rates (dilution rates) and carbohydrate composition (experiments) are represented by unique shapes and colors, respectively.

Metabolomics data is a sensitive and robust method for assessing community-level dynamics and provides a holistic overview of community-level metabolic biotransformations in response to changes in carbohydrate sources and resource availability. Figure 7 represents the targeted metabolic fingerprints for each carbohydrate source and feeding rate and shows the highly reproducible dynamics from two independent experiments and triplicate biological replicates per feeding rate. Lactate concentrations increased across decreasing retention times when sucrose, lactose, or GOS were substrates; however, the presence of bile salts (GOS+BS, hGOS+BS) resulted in an overall reduction of lactate (Figure 7a). Acetate concentrations followed a parabolic trend across feeding rates, peaking at either 0.08 or 0.12 h^−1, depending on the specific carbohydrate composition (Figure 7b). Moreover, we observed an inverse correlation between acetate and CO₂ levels, which supports previous hypotheses that CO₂ depletion could cause a shift from propionate to acetate production.⁵⁹ Propionate concentrations increased with feeding rates in the lactose, GOS, GOS+BS, and hGOS+BS conditions, mimicking the relative abundance of Veillonella spp. (Figure 7c and Table S1), a known propionate producer.⁶⁰

Figure 7.

Targeted metabolite analysis from communities cultured in five different dietary carbohydrate compositions (lactose, sucrose, GOS, GOS + BS, hGOS + BS). Each feeding rate tested shows average steady-state CO₂, H₂, lactate, and targeted short-chain fatty acids (SCFA); acetate, propionate, and butyrate were obtained from two independent experiments for each dietary carbohydrate source. Panels A, B, C, and D display the average values of lactate, acetate, propionate, and butyrate concentrations in g L⁻¹, respectively. Panels E and F display an average of 400 points in the steady state collected by online monitoring of CO₂ and H₂ generation. The L_A term in the productivity units is calculated and dependent on a reactor active volume of 350 mL. In each boxplot, a line represents the median, while the mean is represented by a small square. Pairwise statistics performed for each condition’s metabolic profile is presented in table S2.

Interestingly, the highest levels of butyrate were observed with sucrose as a substrate followed by GOS. Both conditions exhibited statistically significant differences in butyrate levels compared to GOS+BS and hGOS+BS (p < 5e-5; Table S2). Moreover, the overall metabolic profile of the sucrose condition differed significantly from the hGOS+BS condition across all metabolites except lactate concentrations (p = 0.109). Additionally, the simple carbohydrate substrate conditions containing lactose and sucrose exhibited similar metabolic profiles with only propionate and butyrate levels having statistically significant differences.

The highest value of CO₂ was observed with sucrose as a carbon source. When bile salts were present, we observed a marginal decrease in CO₂ rates (Figure 7e). Conversely, the conditions where bile salts were added to GOS and hGOS showed increased H₂ production at high feeding rates (Figure 7f). The increased H₂ production observed under conditions with bile salts was moderately correlated with an increase in the relative abundance of Bacillota (R² = 0.57229; p < 0.001). The addition of bile salts also consistently and reproducibly impacted butyrate production in GOS+BS and hGOS+BS (Figure 7d and Table S2).

Discussion

In this study, we employed single-step bench-scale PBRs²⁸ to investigate consortium re-assembly and its metabolic compositional stability in response to press disturbances. The systematic execution of our experiments, maintaining a consistent environment and nutrient supply, also involved seeding the PBR with the same pooled stool inoculum across all experiments. Thus, using a PBR as the culturing system and operating it dynamically, we assessed community behavior when cultivated on different carbohydrate compositions and retention times. Retention times were chosen to mimic the physiological ranges of the average human infant (8.3–25 h) and were tested incrementally.

This approach unveiled how each supplied carbohydrate composition reproducibly leads to non-stochastic microbial consortium re-assemblies and constant profiles of its matching metabolic cross-feeding products, including the gases. The data also showed that microbial consortia, when exposed to press disturbances (in this case, different carbohydrate structures, and bile salts), reached increasing cohesivity correlating with the higher complexity of the carbohydrate structures, as demonstrated by the high reproducibility of our data. This suggests that the stochastic element of community assembly associated with in-vivo systems is likely a contribution of the host and its physiology.

Press disturbances caused by increasing feeding rates were applied to the initial stages of re-assembled consortia to provide a glance at community stability under the effect of reduced intestinal transit times. Goldford et al. stated that, in contrast with bottom-up studies with a few species, we need to study the assembly of large numbers of multispecies microbiomes under highly controlled and well-understood conditions to be able to contrast ecological theory with experimental data.⁶¹ Their study with soil samples showed that a similar family-level composition based on experimental conditions arose in their 12 communities tested despite very different starting points. They concluded that the source of carbon supplied governed the community assembly.

In vitro systems eliminate uncontrolled factors from the host intestinal tract but provide an extensive surface area, allowing for local interactions resembling the gut environment. Increasing the surface area allows for the retention of slow-growing and non-proliferating microorganisms. Since PBRs resemble size-exclusion chromatography columns, smaller particles are retained for longer periods in the column, and the increase in carbon availability at faster supply rates provided an opportunity for certain members of the community with less efficient carbon uptake systems to flourish, as shown by increasing values of alpha diversity of the assembled communities. This retention phenomenon was particularly apparent in the experiment utilizing commercial GOS, where only 55% of the carbon exists as difficult-to-digest prebiotics. As the dilution rate was stepwise increased, more lactose and glucose were available to species unable to utilize GOS, allowing for their proliferation. We speculated that the pooled fecal inoculum used for each experiment contained non-proliferating microbes, which were retained in the packed-bed system long enough to germinate when nutrient excess was available in the environment, as demonstrated by the differential abundance data analyses. Previously, Browne et al.⁶² hypothesized that sporulation was an unappreciated basic phenotype of the human intestinal microbiota with a profound impact on microbiota persistence and spread between humans. The authors isolated and purified bacteria representing 66 distinct ethanol-resistant species, representing spore-producers, distributed across 5 known families and 2 newly identified candidate families that culturing methods allowed for the identification of novel species from fecal samples.

Network analysis identified Thomasclavelia ramosa as a central player under the hGOS+BS condition. T. ramosa (previously Clostridium ramosa)⁶³ belongs to the family Coprobacillaceae and, although its role in gut health has not been fully elucidated, it generates a plethora of enzymes that facilitate growth using hGOS as substrate, and can use 44 different substrates including N-acetyl-D-galactosamine, N-acetyl-D-glucosamine, cellobiose, dextrin, and lactose. Moreover, T. ramosa encodes a N-acetyl-β-glucosaminidase, a mucin-degrading enzyme.⁶⁴ Considering LacNAc is a major mucin component, T. ramosa’s centrality in this condition becomes apparent.⁶⁵ Conversely, the sucrose co-occurrence network showed a more stochastic and compositionally variable consortium across feeding rates, suggesting a consortium highly responsive to increases in resource availability and, consequently, a change in composition in response to the nutritional disturbance by sucrose. This is consistent with the Metabolically Cohesive Consortia (MeCoCo) hypothesis proposed by Pascual-García et al.¹⁸ proposing that metabolically cohesive consortia assembled in a non-stochastic way present a more stable phenotype that can resist invasion due to the inability of other species to take advantage of the cross-feeding relationships from established members.

In addition to assessing consortia compositional stability, we evaluated the consistency of the metabolic profiles for each assembled consortia using CO₂, H₂, SCFAs, and lactate metabolomics data. Most of the carbon in our system was recovered as lactate, acetate, propionate, butyrate, and CO₂ with biomass accounting for the remaining few percentage points. Our analysis revealed that GOS as a substrate was associated with lactate production. Lactate, a key player in cross-feeding interactions, serves as an indicator of the proliferation of fast-growing lactate producers (mostly lactic acid bacteria of the genera Lactobacillus and Enterococcus) and the lactate consumers’ inability to utilize it at increased accumulation rates.⁶⁶ Additionally, significantly higher butyrate concentrations were noted when sucrose was the dietary carbohydrate source provided. We previously reported lower levels of butyrate in the stool of mice fed GOS compared to the control chow diet.⁶⁷ At the time, we speculated that the lower amount of fecal butyrate was due to the consumption by colonic cells. In addition, our finding is consistent with mouse studies fed sucrose- and starch-based diets, and an in vitro study conducted in the SHIME system fed with starch, lending further validity to our system.^67,68

Hydrogen metabolism in the human gut is driven by fermentative H₂ production and interspecies H₂ reoxidation, primarily by sulfate-reduction, acetogenesis, and methanogenesis.⁶⁹ A marked increase in H₂ generation was observed in GOS and hGOS in the presence of bile salts. Given that the amount of H₂ found in the lumen of the large intestine is mainly impacted by the rate of production versus the rate of utilization of H₂ by the gut microbiota,⁷⁰ we can speculate that bile salts reduced the abundance of H₂-utilizers (methanogens or sulfate-reducer bacteria). While bile salts were included in this study to increase the physiological relevance of the results obtained, future studies should explore their role as modulators of the microbial community. Insufficient production of H₂ has been linked to Parkinson’s disease, while excess H₂ production has been connected to lactose intolerance.^71,72

While animal model experimentation is still a required prelude to clinical trials, in vitro systems can help reduce the variability observed in vivo. Recent in vitro studies have evaluated various methods of inoculum preparation from natural human fecal consortia to increase reproducibility.^73–75 Here, we showed the role of complex carbohydrates in modulating the reproducible reassembly of a gut-derived consortium and the consistency of its associated metabolic profiles. Our experiments showed that feeding increasingly complex carbohydrate compositions leads to metabolically cohesive reassembled consortia that are also resistant to feeding rate disturbances. Furthermore, the PBR-based platform enabled the retention and cultivation of fastidious and slow-growing gut microbes providing the high surface area characteristic of the gut environment while allowing for the creation of local micro niches.

Limitations of the study

Although continuous flow bioreactor platforms operated dynamically do not capture the complexities of host–microbiome interactions, our methodology enabled the examination of thriving balanced consortia and their dynamic coexistence, including their corresponding metabolic cross-feeding relationships, under highly standardized conditions. This approach provided valuable insights into the taxonomic composition, microbial associations, and ecological succession patterns within the microbial consortia, shedding light on factors like persistence and turnover. However, we recognize that our in vitro model does not fully replicate the intricate interactions between the microbiome and the host, nor does it account for factors such as immune responses and host-specific environmental cues.

The vulnerability of the developing gut ecosystem during infancy underscores the importance of elucidating the mechanisms by which press disturbances can lead to persistent alterations in microbial consortium structure and function, ultimately influencing human health trajectories. Our study used a pooled inoculum from multiple donors to provide sufficient diversity and to create a standardized starting point for evaluating microbial responses to dietary disturbances. The pooling strategy enabled us to overcome the limited amount of individual stool samples to standardize initial conditions and observe the re-assembled community’s response to dietary disturbances while also eliminating the interindividual variation outside of our intent. As a result, the findings cannot be generalized to reflect individual-specific microbiome responses. Future studies could explore how such disturbances affect microbiomes at the individual level, which would provide a more personalized understanding of microbial resilience and stability.

In terms of methodology, the 192-h adaptation period, equivalent to eight volume replacements, effectively dilutes and washes out transient environmental sequences, allowing only actively dividing and metabolically contributing OTUs to persist across feeding rate disturbances in the reassembled community structure. This helped to minimize the impact of spurious sequences on the conclusions drawn from the study. Nevertheless, while the pooling strategy helped to standardize initial conditions, we acknowledge that artificially homogenizing the inoculum is a limitation when considering the broader applicability of the results. Although this was appropriate for our study’s focus on community-level responses, our findings suggest that the resulting community dynamics were robust and reproducible under-controlled conditions, generating consistent metabolic fingerprints.

Additionally, we acknowledge the limitations inherent in using 16S rRNA amplicon sequencing techniques, particularly in the context of Operational Taxonomic Units (OTUs). While the OTU approach was selected to provide a broad overview of community structure and to avoid disproportionate filtering that might omit valuable contextual sequences, it does have limitations, such as reduced resolution at the species level and potential difficulties in distinguishing closely related strains. This may overlook important functional differences between strains within the same species. Although our study focused on community-level dynamics, leveraging metabolic outputs to provide insight, future work could benefit from employing Amplicon Sequence Variants (ASVs).

Funding Statement

This project was supported by the North Carolina State University Department of Plant and Microbial Biology. HDW and MM were supported by the graduate student support plan (GSSP) provided through the Graduate School of North Carolina State University. MM received fellowship support provided by the Genetics and Genomics Academy at North Carolina State University for the 2020–2021 academic year. The UNC Microbiome Core is supported in part by project P30 DK034987 (Center for Gastrointestinal Biology and Disease) and project P30 DK056350 (UNC Nutrition Obesity Research Center).

Disclosure statement

M. Andrea Azcárate-Peril and José M. Bruno-Bárcena co-founded the company UMMINO.The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Data availability statement

The raw sequencing data has been deposited into the National Center for Biotechnology Information (NCBI) provided by the National Institute of Health (NIH) under BioProject accession PRJNA681811.

Contributions

JBB and MAA-P conceptualized, designed, and supervised the project. HW performed the experiments and collected the data. HW, MM, JBB, and MAA-P equally contributed to the data analysis, interpretation, and writing the manuscript. RK contributed to the fecal material used for this project.

Ethics declarations

Ethical approval for this study was provided by the University of North Carolina at Chapel Hill under IRB# 12–0809.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/19490976.2024.2440120