Dietary fiber monosaccharide content alters gut microbiome composition and fermentation

ABSTRACT

Members of the mammalian gut microbiota metabolize diverse complex carbohydrates that are not digested by the host, which are collectively labeled “dietary fiber.” While the enzymes and transporters that each strain uses to establish a nutrient niche in the gut are often exquisitely specific, the relationship between carbohydrate structure and microbial ecology is imperfectly understood. The present study takes advantage of recent advances in complex carbohydrate structure determination to test the effects of fiber monosaccharide composition on microbial fermentation. Fifty-five fibers with varied monosaccharide composition were fermented by a pooled feline fecal inoculum in a modified MiniBioReactor array system over a period of 72 hours. The content of the monosaccharides glucose and xylose was significantly associated with the reduction of pH during fermentation, which was also predictable from the concentrations of the short-chain fatty acids lactic acid, propionic acid, and the signaling molecule indole-3-acetic acid. Microbiome diversity and composition were also predictable from monosaccharide content and SCFA concentration. In particular, the concentrations of lactic acid and propionic acid correlated with final alpha diversity and were significantly associated with the relative abundance of several of the genera, including Lactobacillus and Dubosiella. Our results suggest that monosaccharide composition offers a generalizable method to compare any dietary fiber of interest and uncover links between diet, gut microbiota, and metabolite production.

IMPORTANCE

The survival of a microbial species in the gut depends on the availability of the nutrients necessary for that species to survive. Carbohydrates in the form of non-host digestible fiber are of particular importance, and the set of genes possessed by each species for carbohydrate consumption can vary considerably. Here, differences in the monosaccharides that are the building blocks of fiber are considered for their impact on both the survival of different species of microbes and on the levels of microbial fermentation products produced. This work demonstrates that foods with similar monosaccharide content will have consistent effects on the survival of microbial species and on the production of microbial fermentation products.

INTRODUCTION

The human diet primarily comprises fats, proteins, and carbohydrates, with the latter macronutrient contributing the majority of calories in most societies (1). Besides starches and sugars that are readily digested by mammals, dietary carbohydrate sources include various types of indigestible fiber (2). Though molecules such as cellulose are relatively inert, many other types of dietary fiber act as “microbiota-accessible carbohydrates” (MACs), which serve as substrates for metabolism by the gut microbiota (3). These carbohydrates are fermented into bioactive end products such as short-chain fatty acids (SCFAs) that may influence host physiology by altering the pH of the intestinal lumen, contributing to energy balance, and regulating transcription in a variety of metabolic and immune cells by signaling via the receptors Ffar2 and Ffar3 (1). While many dietary plants, including beets, alfalfa, oats, corn, and soy, are recognized as potential sources of MACs (4), the specific chemical structures within these carbohydrates that facilitate microbial metabolism are often unknown.

Knowledge of diet-derived MAC structure is important because distinct carbohydrates are known to influence the composition and function of the mammalian gut microbiota. Many strains of gut bacteria are predicted to encode specific carbohydrate-binding molecules alongside hundreds of carbohydrate-active enzymes that may recognize specific linkages (5). These enzymes’ mechanisms of action may depend on glycan complexity (6) and can entail many coordinated steps (2, 3). In some cases, the metabolism of a carbohydrate may be performed by multiple species (7, 8). Gut bacterial genera such as Bacteroides and Prevotella express diverse glycoside hydrolases, polysaccharide lyases, and carbohydrate esterases, enzymes that are known or predicted by proteomics to break down distinct polysaccharides, such as xylans, pectin, and arabinogalactans (9, 10). The expression of carbohydrate-active enzymes varies extensively within genera and species (11); for example, the species Bifidobacterium longum includes the infant-associated B. longum subsp. infantis, which encodes a gene cluster for metabolizing human milk oligosaccharides (HMOs) (12), as well as adult-associated strains of B. longum subsp. longum that specializes in plant oligosaccharide metabolism (13). Certain microbes can be enriched by targeted provision of specific carbohydrate structures, such as acetylated galactoglucomannans (14, 15) and arabino-oligosaccharides (4). Species of the genus Bacteroides are known to compete for multiple glycans, such as arabinan (16), but may coexist due to differences in physicochemical complexity (17) or substrate prioritization (18), illustrating the intimate relationship between carbohydrate structure and microbial ecology. These bacterial-fiber relationships can hold true across host species, which is why humans, canines, and felines have generally similar gut microbial phylogeny and functional capacity (19). Because cats have reduced digestive enzyme capacity (20), cats make a good model organism to test for associations between fiber structure and microbiome response as a greater percentage of the carbohydrate in food might be available to the gut microbes.

Despite increasing recognition of carbohydrate structure as a determinant of gut microbial dynamics, the carbohydrate content of many foods remains poorly characterized, and the mechanisms of gut microbiome metabolism are only partially elucidated. Starchy plants are staple foods in many human diets, but dietary starch encompasses both soluble amylopectin that is readily digested and five types of resistant starch, which may be microbiota-accessible (21). MACs from different food sources may differ in their discrete structure in terms of monosaccharide content, chemical linkages, isomerism, chain length, and modifications such as methylation and sulfation, all of which may differentially affect their degradation by specific gut microbial enzymes (22, 23). Fermentation by the gut microbiota can be impacted by relatively fine differences in MAC structure, such as arabinoxylans from different kinds of wheat bran (24, 25), the degree of polymerization of oligosaccharides derived from sugar beet pectin (26), and the linkages found in glucans (27). However, due to the complex physicochemical properties of food carbohydrates (28) and the analytical difficulties of isolating and analyzing glycans (29), the exact composition of most diet-derived MACs is unknown.

Recently, novel approaches based on ultra-performance liquid chromatography and triple-quadrupole mass spectrometry (UPLC-QqQ-MS) are able to address the challenges of studying the fine structure of complex carbohydrates (30, 31). These techniques can reveal, in high throughput, the monosaccharide composition of any carbohydrate of interest (30), data that have been collected for more than 800 foods in the “Davis Food Glycopedia” (31). Precise monosaccharide composition offers a simple, generalizable means of comparison between numerous dietary fibers, permitting the analysis of how dietary carbohydrate structure affects gut microbiome function at an unprecedented scale.

The present study aimed to determine if differences in fiber monosaccharide content influenced the composition and metabolic function of a model gut microbiota. Data on the monosaccharide composition of foods, portions of which have been published in the Davis Food Glycopedia, guided the selection of 55 qualitatively diverse fibers with varied monosaccharide content. In order to model mammalian microbiome function in vitro, each fiber was fermented over 72 hours in a bioreactor. 16S rRNA gene sequencing was performed in order to identify changes in the microbiome in tandem with the quantification of pH and short-chain fatty acids.

MATERIALS AND METHODS

Monosaccharide composition

Fiber monosaccharide composition data were obtained with UPLC-QqQ-MS, as previously described (30).

Food processing

Foods (Table S1) were processed to model mastication prior to digestion. All foods were purchased from local grocers and stored at 4°C prior to processing in the Food Instruction Lab at the Robert Mondavi Institute for Wine and Food Science. Foods with low water content, such as fried seaweed, were ground into a powder using a Vitamix. Foods with high water content, such as fresh apples, were diced and ground into a slurry using a Vitamix. In some cases, water was added to achieve a homogeneous slurry. The addition of water was recorded and accounted for in the in vitro digestion calculations. After processing, foods were stored in Quart Ziplock Bags at −20°C until digestion.

In vitro digestion of foods

The in vitro protocol to mimic cat digestion was modeled on the INFOGEST protocol, a previously validated human in vitro digestion system (5). The following changes were made to adapt the protocol to the cat digestive system. Simulated salivary fluids (SSFs), simulated gastric fluids (SGFs), and simulated intestinal fluids (SIFs) were adapted from the INFOGEST protocol to match the minerals found in the cat saliva, gastric fluid, and intestinal fluid, respectively. In cases where a range of mineral concentrations is common, the average between the minimum and maximum reported values was used. Table S2 details the range of mineral concentrations reported as well as the reference material. Table S3 details the final composition of 1.25X SSF, SGF, and SIF used.

In the oral phase, salivary alpha-amylase was removed, given that cats do not secrete the enzyme, and pH was adjusted to 7.5 based on the average pH of cat saliva (32). The food slurry was mixed with 1.25× SSF, pH of the mixture was adjusted to 7.5, and the final volume was adjusted using deionized (DI) water for a final ratio of 1:1 food slurry and SSF. The oral bolus was incubated at 37°C with agitation for 2 min. To simulate gastric digestion, the oral bolus was mixed with 1.25× SGF, 2,000 U/mL pepsin (Sigma P6887), and adjusted to pH 2.5 using 6 M HCl (33). DI water was added for a final ratio of 1:1 oral bolus and SGF. The gastric fluid was incubated at 37°C with agitation for 3 hours. While gastric emptying time varies by the individual cat and by the type of food eaten, 3 hours was chosen as a standard based on the average gastric emptying time for dry cat food (34). To simulate intestinal digestion, the gastric chyme was mixed with 1.25× SIF, pancreatin based on trypsin activity at 100 TAME Units/mL (Sigma P7545) (35), 11.5 U/mL amylase (Sigma A6814), and 5 mg/mL bovine bile (Sigma B3883) (32). The contribution of amylase in cats is very low, given that their diets are rich in protein. However, there is some evidence that amylase secretion in the pancreas increases after prolonged feeding of a high-carbohydrate diet. The amylase concentration was chosen based on the average amylase secretion from four healthy adult cats fed a diet containing potato starch for 3 weeks (36). The pH of the mixture was adjusted to 6.0 using 1 M NaOH (33). The intestinal slurry was incubated at 37°C with agitation for 2.5 hours, a time chosen based on the average intestinal emptying time of the cat (32). After incubation, the intestinal slurry was immediately frozen at −20°C to halt enzymatic activity and stored until further use.

Dialysis and anthrone assays

Intestinal slurries were dialyzed to remove monosaccharides and excess salts using Biotech Cellulose Ester Dialysis Membranes (31 mm flat width, 100–500 D molecular weight cut-off, Catalog No. 888–10723, Repligen, Waltham, MA) Dialysis membranes were cut to a length of 400–550 mm and stored in 1% (vol/vol) formaldehyde (refreshed every 2–3 rounds of dialysis). Prior to each use, dialysis membranes were hydrated and rinsed with distilled water. Approximately 110–220 mL of thawed intestinal slurries were pipetted into each dialysis membrane. The filled dialysis membranes were submerged in a 30 L tub containing 22.5 L of distilled water. No more than 14 dialysis tubes (representing ~350 mL of intestinal slurry) were placed into the same dialysis container. Foods were often dialyzed together, and groupings were chosen to avoid intestinal slurries with either a high volume, high monosaccharide content, or high salt content being dialyzed together. The dialysis container was stored at 2°C for the duration of dialysis. Twice per day, the water was exchanged with fresh distilled water. After 96 hours, the dialyzed intestinal slurries were harvested from the dialysis membranes, placed into Gallon Ziploc Freezer Bags, and stored at −20°C until freeze drying.

Dialyzed intestinal slurries were lyophilized to remove water content using The Scientific Freeze Dryer (4-tray dryer, Medium, Harvest Right, North Salt Lake, UT). The slurries were thawed in a room temperature water bath, and then poured into even layers onto freeze dryer trays. Large volumes (>1.5 L) of dialyzed intestinal slurry were freeze dried in multiple batches and homogenized. Freeze drying conditions were −30°F (−34°C) for 4 hours, vacuum off; 35°F (1.7°C) for 3 hours, vacuum on; 80°F (26.7°C) for 12 hours, vacuum on; 120°F (48.9°C) for 5 hours, vacuum on. The material was then held at 120°F (48.9°C) with the vacuum on until it was retrieved from the machine. The lyophilized powder was then homogenized in a coffee grinder and stored in 50 mL conical tubes kept at −20°C. All lyophilized powders were stored in sealed secondary containers with desiccant prior to anthrone assays.

Anthrone assays were performed to determine the total fiber content of the lyophilized powders. A serial dilution of fibers and amylopectin standards were prepared at 0.75 mg/mL, 0.6 mg/mL, 0.5 mg/mL, 0.4 mg/mL, 0.25 mg/mL, and 0.1 mg/mL. Seventy microliter of serially diluted fibers or amylopectin was combined with 140 µL of 2 mg/mL anthrone in 6 M sulfuric acid in 0.5 mL strip tubes. Reactions were run in duplicate. The strip tubes were centrifuged and run in a Thermocycler at 90°C for 11 min followed by 20°C for 8 min. Samples were transferred to 96-well plates, and fluorescence data were obtained with a plate reader. The total fiber content of the lyophilized powder was then calculated.

Batch fermentation media

To make base media, in 80% total volume of DI water, the following components were dissolved: 0.05 g/L bile salts, 1 g/L casitone (hydrolyzed casein), 0.67 g/L proteose peptone no. 3, 4.83 g/L NaCl, 5 g/L K₂HPO₄, 3.19 g/L KH₂PO₄, 1.93 g/L NaHCO₃, 2.33 g/L Na₂CO₃, 9.76 g/L 2-(N-morpholino)ethanesulfonic acid, 0.50 g/L cysteine-HCl, 0.60 mL/L of 0.1% resazurin in water, 1 mL/L of 500 mg/mL MgSO₄*7H₂O in water, 10 mL/L hemin solution (1 mg/mL in 0.02M NaOH), 2 mL/L Tween 80, 1.80 mL FeSO₄*7H₂O solution (2.78 mg/mL in water and 6 M HCl dropwise to dissolve), 2 mL/L mineral solution (0.5 g/L EDTA, 0.2 g/L FeSO₄*7H₂O, 0.01 g/L ZnSO₄*7H₂O, 0.003 g/L MnCl₂*7H₂O, 0.03 g/L boric acid, 0.02 g/L CoCl₂*6H₂O, 0.001 g/L CuCl₂*2H₂O, 0.002 g/L NiCl₂*6H₂O, 0.003 g/L NaMOO₄*4H₂O dissolved in water), 5.6 mL/L vitamin solution [0.25 g/L menadione, 0.5 g/L biotin, 0.5 g/L pantothenate, 2.5 g/L nicotinamide (niacin), 0.125 g/L vitamin B12, 1 g/L thiamine, and 1.25 g/L p-aminobenzoic acid dissolved in 63% ethanol], and 1 mL of 20 mg/mL CaCl₂*2H₂O in water. This solution was filter-sterilized (0.2 µm pore-size), covered with foil, and placed immediately under anaerobic conditions for 48 hours before use to remove oxygen. Media was made fresh before each experiment.

Batch fermentations of fibers

Batch fecal fermentations were conducted in triplicate with 55 unique fibers, divided into nine experiments. In addition to the fibers, one no carbohydrate control was included in each experiment. Fermentations were conducted in an anaerobic chamber using a MiniBioReactor Array system (37, 38) in batch mode. Feline fecal samples (provided by Mars, Inc., McLean, VA) were used to inoculate fermentations. Because the fecal microbiome of cats has been found to cluster by both subject and the date of sampling within subjects (39), a single inoculum for all fermentations was created by pooling 10 fecal samples with distinct microbial communities. Samples to be pooled were selected based on the microbial communities’ dissimilarities between available samples. Dissimilarities were calculated using 16S rRNA gene sequencing data of each individual available fecal sample and Bray Curtis index (data not shown). The 10 most distinct samples were pulled to maximize microbial diversity and functional genetic richness. The fecal inoculum for each fermentation was placed in the anaerobic chamber, and 25% m/v fecal slurry was made using previously reduced 1× PBS. Slurries were homogenized by vortexing for 5 min and centrifuged at 200 g for 5 min to separate large particles. The supernatant was mixed with fermentation media in a 20:80 ratio and incubated for 72 hours with constant stirring, under anaerobic conditions. Experimental fermentation media contained 1% m/v of the fiber of interest. No carbohydrate was added to control fermentations. One thousand five hundred microliters of samples was collected and split into 0.5 mL and 1 mL aliquots after 0, 24, 48, and 72 hours of batch fermentation. From the 0.5 mL sample, pH was immediately measured as samples were retrieved, and all samples were then preserved at −80°C prior to subsequent analyses (quantification of short-chain fatty acids and 16S rRNA gene sequencing).

Quantification of SCFAs

Sample preparation and derivatization

The SCFA quantification workflow was adapted from a previous publication (40). A pooled standard solution comprising 19 carboxylic acid metabolites was formulated in methanol (MeOH) and serially diluted to concentrations ranging from 0.001 µg/mL to 500 µg/mL, depending on the specific analytes. The batch fermentation supernatants were thawed and centrifuged at 13,500 rpm for 5 min. The supernatant was then collected and diluted 25-fold in MeOH. An internal standard mixture containing 100 µg/mL of d4-acetic acid, 50 µg/mL of d2-indolepropionic acid, and 10 µg/mL of 2-ethylbutyric acid was spiked into all standards and samples at the ratio of 1:10 (vol/vol). Two hundred microliter of acetonitrile (ACN) and 100 µL of derivatization reagent containing 20 mM triphenylphosphine, 20 mM dipyridyl disulfide, and 20 mM 2-picolylamine in ACN was plated in a 1 mL 96-well plate in advance. Ten microliter aliquots of standard or sample was dispensed into each well of the plate, which was then sealed and incubated at 60°C for 10 min. The entirety of the procedure was conducted within a 4°C cold room to minimize the evaporation of volatile analytes. Upon completion of the reaction, the derivatized samples were subjected to drying in a miVac concentrator. The dried samples were then reconstituted with 50% MeOH prior to instrumental analysis. Data are reported in micrograms per milliliter (μg/mL). The analysis was limited to the following metabolites: acetic acid, propionic acid, butyric acid, lactic acid, indole-3-propionic acid (I3P), indole-3-butyric acid (I3B), and indole-3-acetic acid (I3A).

Liquid chromatograpy with tandem mass spectrometry (LC-MS/MS) analysis

Derivatized analytes were analyzed on an Agilent 6495B QqQ MS coupled to an Agilent 1290 Infinity II UHPLC, as previously described in greater detail (40). Separation was performed on an Agilent Poroshell 120 EC-C18 column (2.1 mm × 100 mm, 1.9 µm particle size). Aqueous mobile phase A consisted of 100% nano-pure water. Organic mobile phase B consisted of a 1:1 (vol/vol) ACN/isopropyl alcohol mixture. The following binary gradient was used: 0.00–1.00 min, 5.00% B; 1.00–10.00 min, 5.00%–20.00% B; 10.00–11.00 min, 20.00% B; 11.00–15.00 min, 20.00%–60.00% B; 15.00–16.00 min, 60.00%–5.00% B. One microliter of sample was injected in each run. The mobile phase flow rate was 0.45 mL/min, and the column temperature was set to 45°C. The Jet Stream Technology (AJS) electrospray ionization ion source was employed in positive ion mode, with the following parameters: capillary voltage = 1,800 V, nozzle voltage = 1,500 V, gas temperature = 240°C, gas flow = 20 L/min, nebulizer pressure = 25 psi, sheath gas temperature = 300°C, and sheath gas flow = 9 L/min. Mass spectrometry data were acquired using dynamic multiple reaction monitoring mode.

DNA extraction, library prep, and 16S rRNA gene sequencing

Genomic DNA was extracted using Zymobiomics 96 MagBead DNA Kit (Zymo Research, Irvine, CA) with a Kingfisher Flex automated extraction instrument (Thermo Fisher Scientific, Waltham, MA). As previously described (41), the V4 region of the 16S rRNA gene was amplified in triplicate with barcoded PCR primers F515 and R806. Amplicons were verified by gel electrophoresis, combined and purified, and sent to the UC Davis Genome Center for library preparation and high throughput 250 paired-end sequencing using an Illumina MiSeq. The sequencing run was performed in two batches. A ZymoBIOMICS mock community was used as a positive control.

Bioinformatics

Raw sequencing data were demultiplexed with Sabre (42). Demultiplexed data were imported into QIIME2-2021.2 (43), where data were quality filtered, and reads were processed with DADA2 (44). After filtering, taxonomy was assigned to reads using a pre-generated naïve Bayes classifier trained on Silva 138 99% operational taxonomic units (OTUs) from the 515F/806R region of sequences (45–47), accessed via QIIME2-2021.2 (43). Reads mapped to the microbial orders Caldalkalibacillales, Rhizobiales, Sphingomonadales, Micrococcales, and Bacilliales were present in three negative controls on one DNA extraction plate; therefore, contamination was suspected. Contaminant sequences were statistically identified using decontam (v. 1.1.1) (48) with default settings and pruned from the data set using phyloseq (v. 1.34.0) (49), resulting in the removal of 1/1,854 taxa from batch 1 samples and 21/1,854 taxa from batch 2 samples.

Statistical methods

All statistical analyses were completed in R 4.2.0 statistical software (50) and QIIME2-2021.2 (43).

Density-based spatial clustering of applications with noise

Fibers were clustered based on unadjusted monosaccharide composition using the clustering algorithm DBSCAN (density-based spatial clustering of applications with noise), as implemented in R with the package dbscan (v. 1.1–10) (51). Because the data were high-dimensional and little domain knowledge exists for choosing an appropriate value of the ε parameter (52), hierarchical DBSCAN with simplified hierarchy extraction was performed with a value of minPts = 3.

Data filtering and standardization

Multicollinearity among monosaccharides was identified with caret (v. 6.0–92) with a cutoff of absolute correlation >0.70, which resulted in the monosaccharides fucose, GlcNac, GalNac, and arabinose being filtered from the data set and excluded from subsequent analyses. Linear dependencies and variables with near-zero variance were not detected in monosaccharide data. Values of the retained monosaccharides were multiplied by the mass of fiber (mg) used in fermentations to accurately reflect the amount of carbohydrate present in each fermentation. In order to standardize monosaccharide data across fibers, filtered and corrected monosaccharide values were scaled from 0 to 1 by subtracting the minimum value of each monosaccharide and dividing the result by the difference between the minimum and maximum values for that monosaccharide. The same parameters were used to identify multicollinearity among SCFAs with caret, with the exception that a cutoff of >0.50 absolute correlation was applied, resulting in the exclusion of acetic acid, butyric acid, and indole-3-propionic acid from subsequent analyses. The remaining SCFAs (propionic acid, lactic acid, indole-3-butyric acid, and indole-3-acetic acid) were also scaled from 0 to 1 using the same scaling function as for monosaccharides.

Hierarchical clustering

Fibers were grouped by hierarchical clustering based on filtered, corrected, and scaled monosaccharide composition data. A distance matrix based on Euclidean distance between monosaccharide values was calculated. Hierarchical clustering was performed with the hclust function in the base R package using both complete and average linkage, and Dunn’s index was calculated for different values of k to find its maximum value. The heatmap function was used to plot a heatmap of monosaccharide hierarchical clustering results based on complete linkage. Similarly, filtered and scaled SCFA data were grouped with hierarchical clustering and visualized with a heatmap using the same parameters as for monosaccharides; however, average values of SCFA concentrations were used due to the variation between samples.

Modeling with final pH as outcome

To assess the effect of monosaccharide composition on final pH, we used a generalized linear model as implemented in the base package of R. Only monosaccharides found to have an absolute correlation below 0.70 with caret were eligible for inclusion. Purposeful selection was used to select the variables to include in the final model. Log or Box–Cox transformations were applied to monosaccharide data that deviated from normality, as assessed with the Shapiro–Wilk test. If a monosaccharide had at least one zero value for food, a pseudocount was added prior to transformation.

Alpha diversity

After excluding statistically identified contaminants, samples were rarefied without replacement to a depth of 2,509 reads using the vegan package (v. 2.5.7) (53), which retained 1,881,750 features (15.56% of features) in 750 samples (92.59% of samples). Rarefied data were used to calculate alpha diversity with the Shannon index using the vegan package [vegan package (v. 2.6–2) (53)] with default settings. To assess the effect of monosaccharide composition on longitudinal alpha diversity, a linear mixed effects (LMEs) model was implemented with the lmer function in the package lmerTest (v. 3.1.3) (54). Bioreactor ID was used as the grouping variable. Alpha diversity was observed to vary at baseline, so baseline alpha diversity was considered as a variable in the model. Day, baseline alpha diversity, and each monosaccharide were considered as fixed effects. Bioreactor ID was considered as a random effect since it was expected that the baseline values and rate of change would differ between fermentations. Log or Box–Cox transformations were considered for variables that deviated from normality, as assessed with the Shapiro–Wilk test. However, these transformations did not improve the normality of monosaccharide data across the entire time course of fermentation, so they were ultimately not applied. A modified form of purposeful selection was used to select the variables to include in the final model. Separate models were evaluated for each monosaccharide and baseline alpha diversity; each of these models included day as a fixed effect and fermenter as a grouping variable. Restricted maximum likelihood was set to FALSE since the models compared different fixed effects. Variables that were significantly associated with alpha diversity were then assessed in a combined model, and the assumptions of linearity, homogeneity of variance, and normality of residuals were tested. Although a significant result was obtained in the analysis of variance of the squared residuals for bioreactor ID (P-value = 5.374e-12 by Levene’s test), the histogram of the residuals did not appear bimodal. Therefore, since a q-q plot showed relatively linear residuals apart from extrema, the results of Levene’s test were disregarded, and other variables were added back into the model by purposeful selection. A confidence interval was calculated using the confint function in the base R package.

Regression trees

Regression trees were constructed with the rpart package (v. 4.1.16). For final pH, data were divided into training (131 samples, mean final pH = 6.12) and test (30 samples, mean final pH = 6.15) sets, and the tree was pruned at a complexity parameter of 0.060944. For final alpha diversity, data were divided into training (118 samples, mean final alpha diversity = 2.75) and test (40 samples, mean final alpha diversity = 2.77) sets, and the tree was pruned at a complexity parameter of 0.031896. The performance of trees on test data was summarized with caret.

Beta diversity

Rarefied data were used to analyze beta diversity (between-sample diversity) because read depth varied more than 10-fold between samples. A distance matrix based on Bray–Curtis distance between samples was calculated based on decontaminated 16S rRNA gene sequencing data with the function vegdist in the vegan package (v. 2.5.7) (53). Bray–Curtis distance was selected as a distance metric instead of weighted or unweighted Unifrac (55) due to the presence of unassigned reads that became outliers if phylogeny was taken into account. Sequencing data were subsetted by day with phyloseq (v. 1.34.0) (49). Differences in beta diversity by monosaccharide composition were tested with permutational analysis of variance (PERMANOVA), as implemented by the adonis2 command in vegan, which can accommodate continuous variables. A backward elimination approach was used to select the final variables for each day’s model. Ordination with nonmetric multidimensional scaling (NMDS) was performed with the metaMDS command in vegan to visualize day 3 beta diversity data. Ordinations with points colored using a gradient based on fiber monosaccharide content and pH were created with the plotfunctions package (v. 1.4).

Differential abundance

Differential abundance based on monosaccharide composition and SCFA concentration was tested with Analysis of Composition of Microbiomes (ANCOM) (56) in QIIME2-2021.2. Because ANCOM requires discrete variables, monosaccharide composition and SCFA concentration values were classified into tertiles with the functions quantile and cut in the base R package. Tertiles were chosen over quartiles or quintiles due to the relatively large number of very small values in many variables. 16S rRNA sequencing data were filtered to exclude samples prior to day 3. Taxonomy was collapsed to the genus level. Pseudocounts were added due to the presence of zeros in microbiome data. ANCOM was run separately for each monosaccharide and SCFA tertile.

RESULTS

In order to test the effects of fiber monosaccharide composition on the outcome of gut microbial fermentation, dietary fibers from 55 unique foods with known monosaccharide composition were chosen (31). These food fibers were selected to represent a diverse range of dietary fruits, vegetables, and starches common in human diets. The relatively large number of fibers analyzed presented a challenge for robust cross-fiber comparisons. We, therefore, sought to reduce the number of dimensions with clustering analysis. Results of clustering with hierarchical DBSCAN (Fig. S1) identified few clusters based on monosaccharide composition with unbalanced membership and a high degree of noise. Moreover, hierarchical clustering identified a large number of clusters (data not shown), with Dunn’s Index maximized at k = 43. Therefore, instead of grouping fibers into clusters, monosaccharide composition was compared directly across fibers.

The monosaccharide composition of each fiber was documented in terms of the content of glucose, galactose, fructose, xylose, arabinose, fucose, rhamnose, glucuronic acid (GlcA), galacturonic acid (GalA), N-acetylglucosamine (GlcNac), N-acetylgalactosamine (GalNac), mannose, allose, and ribose (Table S1). Some sugars, such as ribose and fucose, were relatively scarce in most food fibers in this data set, while others, such as glucose, were present in many fibers. Some monosaccharides were found to exhibit multicollinearity, and the sugars fucose, GlcNac, GalNac, and arabinose (absolute correlation > 0.50) were filtered from the data set. While the data for some monosaccharides were zero-inflated, no monosaccharides were found to have near-zero variance (data not shown).

In order to determine how fiber monosaccharide composition affected fermentation, each fiber was batch fermented in triplicate for a time course of 72 hours in the presence of a pooled feline fecal inoculum. Samples included in the inoculum were chosen to maximize the taxonomic richness and the functional capability to utilize different fiber substrates (see details in the Materials and Methods section). Prior to fermentation, foods were processed to model mastication and exposed to simulated digestive fluids in vitro to mimic feline digestion, after which free monosaccharides were removed with dialysis, and the total fiber content was measured with anthrone assays (see Materials and Methods) to create a standard concentration of 1% m/v of the fiber of interest in each fermentation. For most fibers, pH decreased over the course of the experiment (Fig. 1), indicating that fermentation generally occurred regardless of fiber source. However, the mean final pH ranged from 5.14 for kabocha squash flesh to 6.75 for radish bulb, suggesting that this pooled microbial community could not fully ferment all fibers. In order to determine if fiber monosaccharide composition contributed to the observed variance in the extent of fermentation, we modeled the effect of monosaccharide content on the final pH with generalized linear regression (Table 1). The resulting model found that glucose content was significantly associated with a lower final pH, while xylose content was significantly associated with a higher final pH. Initial pH was not predictive of final pH, suggesting that monosaccharide composition was an important determinant of fermentation. Since glucose was the predominant sugar in most fibers, we assessed if total monosaccharide content was equally predictive of final pH and found that the more nuanced model was slightly superior [Akaike information criterion (AIC) = 53.821 for the model summarized in Table 1 and AIC = 54.969 for the model based on total monosaccharide content].

Fig 1.

pH change over the course of 3 days (72 hours) food fiber fermentations. pH was measured every day (24 hours). Lines show mean ±2 SD. Food fibers were processed as described Materials and methods.

TABLE 1.

Generalized linear model to predict final pH from fiber monosaccharide composition^a

Variable	Transformation	Beta-coefficient	95% Cl	P-value
Intercept	NA	6.469	6.196–6.742	<2e-16
Glucose	Box-Cox	−0.0216	−0.0407 to −0.00260	0.0267
Xylose	Box-Cox	0.125	0.0200–0.230	0.0206
Rhamnose	Box-Cox	−0.656	−1.509 to −0.197	0.129

^a The influence of each variable on final pH was assessed with purposeful selection modeling. Box-Cox transformations were applied to monosaccharide data due to deviations from normality. Rhamnose was included in the model as a confounder of xylose. Monosaccharide data were scaled by minimum and maximum values.

Since monosaccharide content was found to influence the final pH of fiber fermentations, we hypothesized that pH differences resulted from variation in the SCFAs produced by the gut microbiota during the metabolism of distinct carbohydrate structures. After each 72-hour fermentation, the final concentrations of 19 carboxylic acid metabolites were measured by LC-MS/MS after derivatization (see Materials and Methods). We built a regression tree (Fig. 2) to infer the contributions of individual SCFAs to final pH. Comparison of observed and predicted values (r-squared = 0.471, root-mean-square error = 0.311, and mean absolute error = 0.268) supported our hypothesis that specific SCFAs influenced final pH, with lower pH concentrations predicted by higher concentrations of lactic acid, propionic acid, and indole-3-acetic acid.

Fig 2.

Regression tree predicting pH from SCFA levels of fecal fermentations. I3A, indole-3-acetic acid. Color intensity reflects the predicted pH of that node (darker equals a higher pH value).

Given that specific members of the pooled gut microbiota differ in their ability to produce SCFAs, we next sought to identify how the composition of the initial microbial community changed over the course of fermentation in response to variation in monosaccharide composition. Four samples (0, 24, 48, and 72 hours) were taken from each fermentation for 16S rRNA sequencing. The relative abundance of microbial families (Fig. 3) was generally similar for 0-hour samples, suggesting that the initial conditions of each fermentation did not vary considerably. Differences in relative abundance were apparent between fiber sources in 24–72 hour samples. The alpha diversity (richness and evenness) of microbiome samples was quantified with the Shannon index after rarefaction to a depth of 2,509 reads, which retained 1,881,750 features (15.56% of features) in 750 samples (92.59% of samples). A LME model (Table 2) found that baseline alpha diversity was not predictive of longitudinal alpha diversity at 24–72 hours. However, xylose was significantly associated with lower alpha diversity, while rhamnose, GalA, and GlcA were significantly associated with higher alpha diversity, suggesting that the composition of the microbial community changed due to differences in monosaccharide content.

Fig 3.

Average relative abundance of microbial taxa by fiber source at each time point of fecal fermentation. Panel A, Day 0; Panel B, Day 1; Panel C, Day 2; Panel D, Day 3. Fiber sugar monomer compositions are shown in Table S1.

TABLE 2.

Linear mixed effects model to predict longitudinal alpha diversity^a

Variable	Estimate	95% Cl	P-value
Intercept	2.190	2.090–2.290	<2e-16
Day	0.136	0.105–0.166	<2e-16
Xylose	−0.00708	−0.0127 to −0.00143	0.0145
Rhamnose	0.149	0.0941–0.204	2.99e-07
GalA	0.0394	0.0238–0.0550	1.71e-06
GlcA	0.609	0.100–1.118	0.0195

^a The model included baseline alpha diversity and each monosaccharide as a fixed effect and bioreactor ID as a random effect. Modeling was performed with Ime4. Monosaccharide data were normalized by scaling from 0 to 1.

To determine if observed differences in alpha diversity were accompanied by functional differences, we built a regression tree (Fig. 4) to predict the alpha diversity of unknown samples based on the concentration of SCFAs. Comparison of observed and predicted values (r-squared = 0.675, root-mean-square error = 0.306, and mean absolute error = 0.231) indicated that final alpha diversity was predicted by lactic acid and propionic acid, with higher lactic acid concentration associated with a lower final alpha diversity and higher propionic acid concentration associated with higher final alpha diversity.

Fig 4.

Regression tree of association of SCFA with alpha diversity. Color intensity reflects higher predicted alpha diversity.

Since differences in alpha diversity may be accompanied by alterations in the structure of the microbiome, we tested for dissimilarity between microbiome samples from different fermentations (beta-diversity). A distance matrix of Bray–Curtis distances between microbiome samples was calculated. PERMANOVA of the Bray–Curtis distance matrix indicated that the content of the monosaccharides glucose, galactose, fructose, rhamnose, GalA, GalNac, mannose, and allose, as well as sequencing run, were all significantly associated with the dissimilarity of microbiome samples after 72 hours of fermentation (Table 3). Results from 24 and 48 hours were similar to those at 72 hours, with GlcA also a significant variable at 24 hours, while sequencing run was the only significant variable at 0 hours (Table S4). These results suggest that many monosaccharides independently influenced microbial community structure over the course of fermentation. Ordination with NMDS was performed to visualize 72-hour samples according to their Bray–Curtis distances. Ordinations with points colored using a gradient based on fiber monosaccharide content (Fig. 5A) generally agreed with PERMANOVA results, suggesting that samples from the fermentation of fibers with high levels of each monosaccharide were often similar, as measured by Bray-Curtis distance. A gradient plot of final pH (Fig. 5B) suggested that samples with similar final pH levels also tended to be similar, as measured by Bray-Curtis distance.

TABLE 3.

Day 3 PERMANOVA of microbiome data from 72-hour fermentation samples^a

Variable	P-value
Sequencing run	0.003
Glucose	0.015
Galactose	0.033
Fructose	0.013
Rhamnose	0.004
GalA	0.001
GalNac	0.009
Mannose	0.001
Allose	0.016

^a PERMANOVA was performed with 999 permutations on distance matrics of Bray-Curtis distances with the adonis2 function (vegan package). Rarefied 16S rRNA sequence data (rarefraction depth = 2,509) were used due to >10-fold read differences between samples.

Fig 5.

Gradient plots of microbiome samples after 72 hours (day 3) of fermentation. Ordination was performed with NMDS of a matrix of Bray-Curtis distances (two dimensions and stress 0.1525762). (A) Individual monosaccharide component of food fibers used in fermentations. (B) Final pH.

Given differences in microbiome composition and structure were evident between fermentations, we next sought to identify which specific microbial genera responded to differences in fiber monosaccharide content and contributed to the observed final pH differences. Quantitative fiber monosaccharide data were classified into tertiles to test for differentially abundant taxa with ANCOM. ANCOM results at the genus level (Table 4) suggest that differences in the content of 11 monosaccharides were associated with significant differences in the relative abundance of 11 microbial genera, including Lactobacillus, Flavonifractor, and Lachnospira. Notably, the genus Dubosiella was enriched in fermentations of fibers with higher levels of three monosaccharides: allose, GalNac, and GlcA. Higher GalA content was significantly associated with the relative abundance of three genera: Libanicoccus, Alistipes, and Phocea. In order to determine if any of these genera contributed to observed differences in microbiome function, 72-hour fermentations containing different tertiles of SCFAs were also tested for the differential abundance of microbial genera with ANCOM (Table 5). The relative abundance of Lactobacillus, Flavonifractor, Lachnospira, Libanicoccus, Alistipes, and Phocea was significantly associated with the concentration of both lactic acid and propionic acid. Additionally, the relative abundance of Dubosiella was significantly associated with I3B concentration. Our results suggest that variation in fiber monosaccharide content influences both the community composition and metabolic output of a model gut microbiota.

TABLE 4.

ANCOM results for fiber monosaccharide data, including monosaccharide (Sugar), associated taxa (Genus), centered log ratio (CLR), and W statistic^a

Sugar	Genus	CLR	W statistic
Allose	Dubosiella	23.43	97
Fructose	[Eubacterium] brachy group	11.04	54
GalA	Libanicoccus	9.65	81
GalA	Alistipes	8.35	73
GalA	Phocea	9.72	71
Galactose	UVBA1819	10.83	73
GalNac	Dubosiella	57.41	107
GlcA	Dubosiella	23.83	93
Glucose	Holdemanella	8.81	41
Mannose	Clostridium sensu stricto 1	21.6	119
Rhamnose	Lactobacillus	9.54	108
Ribose	Flavonifractor	9.3	67
Xylose	Lachnospira	12.06	118

^a 16S rRNA sequencing data taken after 72 hours of fermentation were tested for differential abundance based on fiber monosaccharide content. Monosaccharide data were classified into tertiles, and sequencing data were analyzed at the genus level.

TABLE 5.

ANCOM results for acidic microbial metabolites (AMMs)^a

AMM	Genus	CLR	W statistic
Propionic acid	Clostridium sensu stricto 1	65.31	118
Propionic acid	Lactobacillus	55.24	118
Propionic acid	Romboutsia	45.88	117
Propionic acid	Alistipes	46.34	113
Propionic acid	[Ruminococcus] torques group	34.57	113
Propionic acid	Colidextribacter	51.11	112
Propionic acid	Unknown genus of family Lachnospiraceae	41.14	111
Propionic acid	Bacteroides	39.72	110
Propionic acid	Peptococcus	29.98	108
Propionic acid	Solobacterium	33.34	108
Propionic acid	[Ruminococcus] gauvreauii group	33.49	107
Propionic acid	Coprococcus	31.29	105
Propionic acid	RF39	28.64	104
Propionic acid	Dialister	28.14	101
Propionic acid	Lachnospira	20.19	100
Propionic acid	Butyricimonas	23.85	99
Propionic acid	Oscillibacter	27.48	99
Propionic acid	Libanicoccus	23.47	99
Propionic acid	Negativibacillus	23.88	98
Propionic acid	Parabacteroides	19.60	97
Propionic acid	Odoribacter	17.77	97
Propionic acid	Enterococcus	16.34	97
Propionic acid	Muribaculaceae	12.46	95
Propionic acid	Fusicatenibacter	12.88	95
Propionic acid	[Clostridium] methylpentosum group	11.73	94
Propionic acid	Sutterella	28.04	94
Propionic acid	Uncultured	14.78	94
Propionic acid	Butyricicoccus	11.24	94
Propionic acid	[Ruminococcus] gnavus group	12.86	93
Propionic acid	Lachnoclostridium	11.07	93
Propionic acid	Barnesiella	6.60	92
Propionic acid	Phocea	11.19	92
Propionic acid	Uncultured genus of family Desulfovibrionaceae	8.81	89
Propionic acid	Flavonifractor	8.36	89
Propionic acid	Fusobacterium	9.32	88
Lactic acid	Clostridium sensu stricto 1	40.70	119
Lactic acid	Lactobacillus	63.10	119
Lactic acid	Romboutsia	47.60	118
Lactic acid	Unknown genus of family Lachnospiraceae	72.04	117
Lactic acid	[Ruminococcus] torques group	63.04	116
Lactic acid	Alistipes	74.12	115
Lactic acid	Bacteroides	76.81	113
Lactic acid	Lachnospira	26.73	109
Lactic acid	Enterococcus	27.72	109
Lactic acid	Colidextribacter	40.27	109
Lactic acid	Subdoligranulum	29.17	108
Lactic acid	[Ruminococcus] gauvreauii group	37.16	107
Lactic acid	Coprococcus	30.29	106
Lactic acid	Solobacterium	34.58	106
Lactic acid	Dialister	27.25	102
Lactic acid	Butyricimonas	47.84	102
Lactic acid	Oscillibacter	38.21	101
Lactic acid	Fusobacterium	22.46	101
Lactic acid	Peptococcus	19.79	100
Lactic acid	Parabacteroides	17.94	100
Lactic acid	Libanicoccus	27.79	99
Lactic acid	Lachnospiraceae	11.09	99
Lactic acid	Uncultured genus of family Desulfovibrionaceae	18.03	98
Lactic acid	UBA1819	9.74	98
Lactic acid	Phocea	17.09	97
Lactic acid	Odoribacter	16.69	97
Lactic acid	Flavonifractor	16.38	96
Lactic acid	Negativibacillus	19.64	95
Lactic acid	[Ruminococcus] gnavus group	14.21	95
Lactic acid	Uncultured	10.28	95
Lactic acid	Sutterella	13.87	94
Lactic acid	Butyricicoccus	9.59	92
Indole-3-butyric acid	Dubosiella	8.30	53
I3A	Mogibacterium	11.01	44
I3A	Collinsella	10.23	39

^a 16S rRNA sequencing data taken after 72 hours of fermentation were tested for differential abundance based on tertiles of each AMM in the corresponding fermentor. Sequencing data were analyzed at the genus level. Only taxa where the null hypothesis is rejected after false discovery rate adjustment (FDR) are shown.

Briefly, we found significant differences in alpha diversity, beta diversity, and microbial community diversity and structure after 72 hours of fermentation based on fiber monosaccharide composition. Final pH was significantly associated with fiber glucose, rhamnose, and xylose content and was predictable from lactic acid and propionic acid concentrations. The relative abundance of several genera, including Lactobacillus and Dubosiella, were significantly associated with both individual fiber monosaccharides and final short-chain fatty acid concentrations.

DISCUSSION

The composition of the colonic microbiota of mammals is shaped in large part by the ingestion of dietary fibers (57), with glycan complexity determining the extent to which individual microbes can process these polysaccharides or their oligosaccharide components (58). Complex dietary glycans are known to nourish extended “food chains” in the mammalian gut, with primary degraders such as Bacteroides spp. encoding polysaccharide utilization loci to recognize and break down specific structures, releasing byproducts that support downstream niches such as secondary degraders and acetogens (59). Dietary fibers were recently shown to vary in the specificity of the changes they induce in microbial community structure, with simple, common fibers like fructooligosaccharides inducing less dramatic and less selective changes than rarer, insoluble fibers such as β-glucan (60). However, though carbohydrates may be studied with a variety of analytical methods, such as ion mobility spectrometry and gas-phase spectroscopy (61), the complexity of food glycans has precluded detailed understanding of their structure. An ultra-high performance liquid chromatography triple quadrupole mass spectrometry approach developed by the Lebrilla and coworkers (30) expands on existing methods in both breadth and depth. Examples of the detailed monosaccharide composition data obtained with this method have been published in the Davis Food Glycopedia (31) for hundreds of common dietary carbohydrate sources, offering relevant structural information for ongoing studies that examine the impact of diet on the human gut microbiota. Similarities or differences in chemical structure may not be intuitively obvious; for example, hierarchical clustering suggested that the monosaccharide compositions of red sweet potato and Japanese sweet potato were relatively distinct, whereas seemingly unrelated food fibers such as buckwheat and radish bulb were similar. These examples suggest that qualitative dietary diversity may not reflect underlying structural diversity, with possible ramifications for studies that seek that examine the gut microbiota’s response to broad dietary patterns, such as the Mediterranean diet (62) or whole-grain consumption (63).

The present study takes advantage of the data in the Davis Food Glycopedia and directly links variation in monosaccharide composition to the outcome of fermentation by the gut microbiome. Modeling revealed that glucose content was the variable most predictive of the final pH of fermentation, a result that likely reflects the wide distribution of glucosidases in the mammalian gut microbiome (64, 65). While a model based on total monosaccharide content performed almost as well as one based on glucose and xylose content, several individual monosaccharides were significantly associated with microbial alpha and beta diversity and community composition, suggesting that total monosaccharide content inadequately reflects gut microbial fermentation. The capacity for metabolism of less abundant monosaccharides may be common across the microbiome but variable within taxa; for example, only certain Bifidobacterium and Bacteroides strains encode sialidases (66), but sialidases and their associated catabolic pathways were found in 80/397 genomes analyzed from the Human Microbiome Project (67). Modeling found that xylose content was significantly associated with a higher final pH and a lower alpha diversity. This result may reflect a relative paucity of enzymes for metabolizing xylose-containing carbohydrates in the microbiome inoculum. Alternatively, it is possible that xylose metabolism is relatively common, but xylose-rich foods support the growth of a narrower range of microbes. This may be due to competition between microbes, leading to reduced net fermentation and alpha diversity. This is plausible because common microbes such as Prevotella and Bacteroides are known to metabolize xylans and arabinoxylans (9, 24, 25), but the fine structure of xylose-containing oligosaccharides may govern antagonistic regulation of bacterial carbohydrate-active enzymes (68) and the ability of secondary degraders to consume xylooligosaccharide breakdown products (58). Further supporting this possibility, a recent study showed that related Prevotella spp. competed for arabinoxylan, with P. intestinalis outcompeting P. rodentium and P. muris (69). A randomized controlled trial showed that arabinoxylan reduced LDL and total cholesterol (70). Additionally, Bacteroides that metabolize arabinoxylans have been shown to release the antioxidant ferulic acid (25, 71), suggesting that gut microbial xylose metabolism may have implications for health. In contrast to xylose, the less abundant sugars rhamnose, GalA, and GlcA were significantly associated with a higher alpha diversity. Structures containing these rarer monosaccharides may support a higher alpha diversity by providing a higher selectivity (17), creating many unfilled ecological niches. Future research should identify carbohydrate-active enzymes that process structures containing rhamnose, GalA, and GlcA to determine which microbes are capable of metabolizing them.

Fermentation pH and microbial diversity were also found to reflect variation in SCFAs, which are known to affect microbial ecology. While changes in dietary fiber consumption are known to rapidly alter SCFA production by the mammalian gut microbiota, such as that of beagles (72), the metabolic fluxes that convert individual dietary monosaccharides into SCFAs are poorly understood. A recent study suggested that SCFAs such as succinic acid, propionic acid, and butyric acid differentially inhibited gut Bacteroidales, but sensitivity to butyric acid depended on the provision of specific carbohydrates (73). This work found that final pH was strongly associated with the levels of acetic acid, possibly because acetic acid is the most widely produced SCFA in the gut [the molar ratio of acetic acid:propionic acid:butyric acid is 60:20:20 (12)]. This finding is consistent with other work, for example, a reduction in mouse cecal pH mediated by the production of acetic acid by Bifidobacterium was shown to protect against infection by Escherichia coli O157:H7 (74). In addition to the association with acetic acid, propionic acid was associated with a lower final pH and with differences in the relative abundance of 35 genera. Using regression trees, higher propionic acid was also predictive of higher alpha diversity, while lactic acid was predictive of lower alpha diversity. While propionic acid is less studied than butyric acid and acetic acid, it was found to increase in human subjects fed corn bran arabinoxylan (75) and has been shown to mediate colonization resistance to Salmonella infection in a mouse model (76), suggesting that future studies should examine its production by genera identified here, such as Lactobacillus and Romboutsia. Propionic acid is produced by multiple routes, including from succinate and via the acrylate pathway (77), possibly explaining why numerous genera were significantly associated with propionic acid levels. While many studies have examined the production of SCFAs by the gut microbiota, the present study also examined a broader range of acidic microbial metabolites. I3A was also found to be significantly associated with a lower final pH, while I3P was significantly associated with the relative abundance of 28 microbial genera, suggesting that it may be produced by a variety of gut bacteria. These indole derivatives are bioactive: I3P signals through the pregnane X receptor to maintain intestinal barrier integrity (78), while I3A signals via the aryl hydrocarbon receptor and may reduce liver inflammation (79). However, the relationship between host diet, microbial metabolism, and the production of these metabolites remains unclear. In the production of acidic microbial metabolites, there is a role for the production of both primary degraders that directly convert food polysaccharides to produce metabolites and “cross-feeders” who produce metabolites by taking up and breaking down byproducts released by primary degraders (80). As an example of cross-feeding, the genome of the butyrate producer Faecalibacterium prausnitzii encodes two loci for catabolizing mannooligosaccharides liberated by primary degradation of β-mannans (81). While our study identified several genera whose relative abundance was significantly associated with monosaccharide content and acidic microbial metabolite levels, such associations may be indirect and depend on interspecific ecological interactions, such as cross-feeding and competition. Future studies should trace the microbial origins of these metabolites in order to establish the foci at which variation in dietary monosaccharide content affects microbial food webs.

The factors that influence both food carbohydrate structure and the ecology of the gut microbiota are multifactorial, so understanding their relationship is challenging. This study suggests that monosaccharide composition provides a generalizable metric for comparing dietary carbohydrates and identifying the functional roles of individual members of the microbiota. The principle of competitive exclusion suggests that each microbe must realize a distinct niche in the gut ecosystem, but this situation may be achieved by multiple possible mechanisms, such as alternative gene regulation and enzymatic specificity (8). Microbes may also exhibit metabolic flexibility; in a study Bacteroides cellulolyticus competed with Bacteroides vulgatus for arabinan, but competition for arabinoxylan between B. celluloyticus and Bacteroides ovatus did not occur because B. ovatus priorritized alternate glycans (16). Despite the complexities of such interactions, monosaccharide composition offers a resource for the large-scale analysis of dietary carbohydrate structure in order to help untangle the mechanisms by which gut bacteria metabolize microbiota-accessible carbohydrates. Our study used a pooled, feline fecal sample with a modified feline artificial digestion process to maximize the microbial accessible carbohydrate content in our experiment, leading to a successful demonstration of the reproducibility of a bioreactor-based system for analyzing the response of a mammalian microbiota to a large set of dietary fibers. This system allows the impact of carbohydrate structure to be examined independently of host factors. We identified little-studied significant associations between monosaccharide content and the relative abundance of specific genera, such as the significant association between higher GalA content and the relative abundance of Libanicoccus, Alistipes, and Phocea. While Phocea and Libanicoccus have been isolated from human stool (82, 83), they are poorly studied and may merit further investigation in relation to diet and metabolism. In contrast, the genus Alistipes is well-known within the human gut for its bile tolerance (84) and ability to perform protein putrefaction (85), but its response to dietary fibers containing GalA has not been studied. Analysis of monosaccharide composition provides a simple way to identify many such potential links between dietary carbohydrate structure and specific gut bacteria. Examples to guide future mechanistic research include past studies that have demonstrated how individual gut bacteria consume HMOs (86), mucins (87, 88), algal porphyrans (7), red seaweed agarose (89), β-mannans (81), and acetylated galactoglucomannans and arabinoglucuronoxylans (15, 90).

Our study thus provides an initial framework for analyzing dietary monosaccharide intake as one of several factors that contribute to the mechanisms of gut microbial metabolism. Future studies may exploit emerging methods, such as chemical and isotopic labeling, in order to trace the fate of individual dietary monosaccharides in single microbial cells (91). While our work identified monosaccharide content as a determinant of fermentation, future studies should examine other facets of carbohydrate structure that are known to affect microbial metabolism, such as linkages (92) and chemical modifications (23, 93). In particular, different glycosyl hydrolases cleave select linkages between monosaccharides, which are critical to defining the microbes capable of consuming a fiber. For example, different microbes are known to have different genetic machinery to digest starches (6) and can enable keystone degraders, who complete external initial cleavage of the starch, to cross-feed secondary degraders (94). As a result, an enhanced understanding of both fiber monomer and linkage content will be critical to understanding microbiome-fiber interaction networks. A recent study supporting this concept compared the human gut microbiome’s response to arabinan-rich pea fiber and homogalacturonan-rich orange fiber in vivo and found that changes in carbohydrate-active enzyme content were negatively correlated with the glycosidic linkages they targeted (95). This argues that monosaccharides should be considered within their broader structural context. More nuanced structural analysis may also use liquid chromatography-mass spectrometry to identify oligosaccharides produced by the cleavage of polysaccharides (96) in order to predict which secondary degraders may be able to exploit the breakdown of dietary glycans. Fiber monosaccharide data may provide input for deep-learning algorithms to predict microbial responses to diet (97). Eventually, knowledge of the intimate relationships between carbohydrate structure and the metabolism of individual microbes may permit the development of more personalized “microbiota-directed foods” (98) that target functional outcomes, such as the production of bioactive microbial metabolites or changes in the host’s plasma proteome (99). A recent human crossover trial found that butyrate production by the gut microbiota in response to the simple glycans inulin, galactooligosaccharides, and dextrin was limited by an individual’s habitual fiber intake (100), suggesting that future interventions may be more successful if they exploit a deeper understanding of dietary fiber structure to identify gaps in the structural diversity of personalized human diets.

DATA AVAILABILITY

Sequencing data from this project are available on NCBI SRA, accession PRJNA1127241.

SUPPLEMENTAL MATERIAL

The following material is available online at https://doi.org/10.1128/aem.00964-24.

Figure S1. aem.00964-24-s0001.tiff.

Hierarchical clustering analysis of 55 food fibers based on absolute monosaccharide composition.

Supplemental tables. aem.00964-24-s0002.pdf.

Tables S1 to S4.

ASM does not own the copyrights to Supplemental Material that may be linked to, or accessed through, an article. The authors have granted ASM a non-exclusive, world-wide license to publish the Supplemental Material files. Please contact the corresponding author directly for reuse.