Metabolic plasticity of the gut microbiome in response to diets differing in glycemic load in a randomized, crossover, controlled feeding study

Abstract

Background

Dietary patterns characterized by low glycemic, minimally processed plant foods are associated with lower risk of several chronic diseases.

Objectives

Evaluate the effects of a low glycemic load (LGL) compared with a high glycemic load (HGL) dietary pattern on stool bacterial community structure and metabolism.

Methods

Participants in this crossover-controlled feeding study were healthy men and women (n = 69). We identified genera, species, and genes and transcripts of metabolic pathways and bacterial enzymes using 16S rRNA gene, metagenomic and metatranscriptomic sequencing, and bioinformatic analysis.

Results

Overall community structure measured by alpha and beta diversity was not significantly different across the diets, although diet did significantly increase the abundance of 13 of 161 genera (P_adj < 0.05) and 5 species in the LGL diet and 7 species in the HGL diet. Gene expression in the hexitol fermentation pathway (β = −1.15; SE = 0.24; 95% confidence interval [CI]: −1.63, −0.67; P_adj = 0.002) was significantly higher in the HGL diet, whereas expression in the L-lysine biosynthesis pathway (β = 0.20; SE = 0.05; 95% CI: 0.09, 0.30; P_adj = 0.03) was enriched in the LGL diet. The beta diversity of expressed carbohydrate-active enzymes (CAZymes) was significantly different between the diets (MiRKAT, P < 0.001). CAZymes enriched in the HGL diet reflected dietary additives, whereas CAZymes enriched in the LGL diet reflected diverse phytochemical intake. There was a significant interaction between homeostasis model assessment of insulin resistance (HOMA-IR) and the coenzyme A biosynthesis I pathway involved in bacterial fatty acid biosynthesis (P_adj = 0.035), which was positive in the HGL diet (b = 0.20; SE = 0.09; 95% CI: 0.02, 0.39) and negative in the LGL diet (β = −0.23; SE = 0.09; 95% CI: −0.40, −0.06).

Conclusions

In healthy humans, diet impacts microbial metabolism and enzymatic activity but not the overall diversity of the gut microbiome. This emphasizes the relevance of dietary components in activating expression of specific bacterial genes and their impact on host metabolism.

This trial was registered at clinicaltrials.gov as NCT00622661.

Introduction

Consumption of a diet with a high glycemic load (HGL), dominated by refined carbohydrates, added sugars, and low-fiber foods, is associated with increased risk of cardiometabolic diseases, several cancers, and mortality [[1], [2], [3], [4], [5]]. In contrast, cohort studies show that consuming a diet with a low glycemic load (LGL), rich in total dietary and cereal fibers, is associated with lower risk of inflammation-based disease linked to excess body fat [[6], [7], [8]]. The measure of GL incorporates both quality (i.e., glycemic index [GI]) and the quantity of dietary carbohydrates. The GI of a food is measured as its impact on postprandial glycemic response relative to a test dose of glucose or white bread. HGL diets dominated by refined carbohydrate-containing foods high in simple sugars are rapidly digested, quickly absorbed, and show a rapid and transient glycemic response [9]. In contrast, LGL diets tend to have higher amounts of soluble and total fiber and an attenuated glycemic response [10].

In a controlled feeding intervention in normoglycemic men and women, we found that metabolic markers responded favorably to an LGL diet pattern compared with an HGL diet pattern. We found lower serum C-reactive protein (CRP) and higher adiponectin among adults with obesity [11], beneficial changes in incretins [12], and reduced postprandial glycemic response [13] with the LGL diet. Metabolomic and lipidomic profiles also differed in response to the dietary intervention with noted differences in microbial metabolites [[14], [15], [16]]. In particular, microbially produced secondary bile acids were positively associated with HOMA-IR overall but not with a specific diet.

Diet impacts the gut microbial community, and long-term dietary patterns are associated with microbial diversity, composition, and gene content [[17], [18], [19], [20], [21], [22], [23], [24], [25], [26]]. Dietary fiber can alter the microbial composition and improve intermediate markers of health outcomes [[27], [28], [29], [30]]; however, many studies to date have focused on adding a single fiber source instead of altering carbohydrate quality within the context of a whole-food diet [[31], [32], [33], [34]]. Interindividual differences in microbial metabolism may impact the magnitude of the health benefits. This analysis aimed to test the effect of LGL and HGL diet patterns on gut bacterial community structure and metabolism in a controlled feeding study in healthy men and women. We show that bacterial carbohydrate-active enzymes (CAZymes) enriched in the HGL diet reflect dietary additives. Higher HOMA-IR was associated with decreased microbial production of vitamin B₅ in an HGL diet.

Methods

Study design

Data and biologic samples for the present analysis were derived from the Carbohydrates and Related Biomarkers (CARB) study, conducted between June 2006 and July 2009 at the Fred Hutchinson Cancer Research Center (FHCC; Fred Hutch). The CARB study was a randomized, controlled, crossover feeding study, the primary aims of which were to evaluate the effects of GL on chronic disease susceptibility biomarkers, e.g., markers of systemic inflammation and insulin resistance, and adipokines [[11], [12], [13]]. Participants were block-randomized, using a randomization procedure developed by the biostatisticians, based on BMI (in kg/m²) and sex, to receive 2 eucalorically similar controlled diets in a computer-generated, randomly assigned order, with a 28-d washout period between the 2 28-d diet periods where participants could eat as desired. The study protocol, including the present ancillary study, was conducted in accordance with the ethical standards of the Declaration of Helsinki and approved by the Fred Hutch Institutional Review Board. All participants gave written informed consent. This trial was registered at clinicaltrials.gov as NCT00622661.

Participants

Details on recruitment have been published previously [11,35]. Briefly, healthy, nonsmoking individuals between the ages of 18 and 45 y were recruited from the Greater Seattle area. Clinic visits were from June 2006 to May 2008. Each participant had 4 clinic visits (other than food pickup visits). Exclusion criteria included impaired fasting glucose measured at a study clinic visit (fasting blood glucose ≥5.6 mmol/L), any physician diagnosed conditions requiring a restricted diet, food allergies, regular use of hormones or anti-inflammatory medication, pregnancy or lactation, or heavy use of alcohol (>2 drinks/d). Participants were entered into the study using a block randomization procedure developed by the biostatisticians in the parent study. Of the 80 participants who completed the parent study, 69 and 64 had complete stool microbiome data available for the 16S rRNA gene and metagenomic/metatranscriptomic analysis, respectively (Table 1; Supplemental Figure 1). At baseline, anthropometric data were collected, including height and weight, and body composition using a whole-body dual-energy X-ray absorptiometry (DXA) scan (GE Lunar DPX-Pro; GE Healthcare), and self-administered questionnaires on demographic characteristics, usual physical activity, and medical history were completed (Table 1).

TABLE 1.

Demographics of participants involved in the dietary intervention¹ with microbiome samples.

Characteristics	Mean (std. dev.)
Sex
Female sex, (%)	34 (49.3)
Age, y	29.4 (8.0)
Race and ethnicity, N (%)
Asian or Asian American	6 (8.7)
Black or African	13 (18.8)
Non-Hispanic White	31 (44.9)
Hispanic	17 (24.6)
Native Hawaiian, Pacific Islander	2 (2.9)
BMI, kg/m2	27.2 (5.7)
Weight, kg	80.3 (21.3)

¹Carbohydrate and Related Biomarkers (CARB) study; 80 participants donated stool samples. Analysis was performed with 69 participants who had full measures. Not included: dropouts (n = 2); missing all or some intervention stools (n = 8) and samples that failed sequencing (n = 1).

Study diets

For the CARB study, we designed 2 diets that were identical in percent energy from carbohydrate (55.0%), protein (30.0%), and fat (15.0%) but differed in GL (125 compared with 250 for the LGL and HGL diets, respectively) and fiber (55 and 28 g/d for the LGL and HGL diets, respectively). A 7-d menu rotation was used for each intervention diet. Each participant completed both diet periods. Based on initial 3-d food record data, physical activity questionnaire, and anthropometrics, each participant was assigned to 1 of 11 energy levels (1800 to 3600 kcal in 200-kcal increments) to maintain their baseline weight. Participants were weighed during every visit to the center, and energy levels were adjusted as needed. The Fred Hutch Prevention Center Shared Resource Human Nutrition Laboratory staff prepared and packaged all the meals for each participant. Participants ate their evening meals in the dining area on weekdays. Breakfast, lunch, and snacks, as well as all weekend meals were packaged and consumed at home. Mean nutrient and food group intakes for the 7-d menu rotation for each of the 11 energy levels for the LGL and HGL diets were calculated by the Fred Hutch Nutrition Assessment Shared Resource using the Nutrition Data System for Research (software version 2024), developed by the Nutrition Coordinating Center, University of Minnesota, Minneapolis, Minnesota, United States.

Blood sampling and analysis

Fasting (12-h) blood draws were collected on days 1 and 28 of each feeding period, processed, and immediately stored at −80°C. Biomarker analyses (HOMA-IR and CRP) were conducted at the University of Washington (Northwest Lipid Research Laboratories and the Diabetes Endocrinology Research Center Immunoassay Core Laboratory) [11,13].

Microbiome sampling

Stool samples were collected prerandomization and at the end of each study period into RNAlater, frozen in a typical home freezer, and delivered to the clinic within 24 h [36]. A stool collection questionnaire, administered at the time of stool collection, collected information on sampling details (date, time, overnight freezing, and any problems). Samples were stored at −80°C until processing. Stool sample aliquots were thawed on ice and pulse homogenized on ice (<30 s) to further disperse the stool sample in RNAlater. Using a wide-bore pipette, samples were aliquoted into 5, 1-mL aliquots and stored at −80°C until analysis. Two 1-mL aliquots of stool in RNAlater were centrifuged at 13 × g to form a pellet, RNAlater was removed, and the remaining pellet was suspended in sterile phosphate-buffered saline and pelleted again. We found previously that extracting 2 aliquots reduces variation in DNA yield and is more representative of the stool sample [37].

Nucleic acid extraction

Two stool aliquots per participant from the end of each intervention period were extracted for DNA following a previously published protocol [38]. Two aliquots of total RNA were extracted, treated with DNase [39,40], and bacterial mRNA was enriched by depleting the 16S rRNA in the stool sample using the Ribominus Bacteria 2.0 kit (ThermoFisher) [37,40,41]. The quality of the remaining RNA (RNA Integrity Number [RIN]) was evaluated using the Agilent 2100 Bioanalyzer (Agilent Technologies).

16S rRNA gene, and metagenomic and metatranscriptomic sequencing

Amplicon sequencing for V_1–3 of the 16S rRNA gene was done as previously reported [42]. Total genomic DNA or rRNA-depleted RNA was shipped to MR DNA for library preparation and metagenomic or metatranscriptomic sequencing on a HiSeq (Illumina) in paired-end format following the manufacturer’s recommendations. For transcriptomics, the rRNA-depleted RNA was converted into double-stranded cDNA using random hexamer primers (MessageAmp TM II; Ambion). The double-stranded cDNA from each sample was then used to prepare sequencing libraries with TruSeq RNA sample preparation kits (Illumina).

Bioinformatics

Stool 16S rRNA gene sequencing bioinformatics analysis was performed as previously reported [42].

Stool metagenomic and metatranscriptomic bioinformatic analysis

Sequence reads were processed for bioinformatic analysis of the metagenomes and metatranscriptome with the KneadData v 0.5.1 quality control (QC) pipeline, which uses Trimmomatic (version 0.36), BMTagger filtering, and decontamination algorithms to remove low-quality read bases and human and other eukaryotic sequences plus the large subunit and small subunit rRNA for metatranscriptomics, respectively [43]. Trimmomatic was analyzed with parameters MAXINFO:80:0.5 and MINLEN:50. Then MetaPhlan2 and HUMAnN2 were analyzed to generate gene and pathway abundances, where pathway abundance was calculated using HUMAnN2’s default pathway gap filling and its harmonic mean over the internal pathway reactions. Functional profiling was performed using HUMAnN2, version 0.11.2 [44] with reads depaired and implementing DIAMOND [45] to map reads against UniRef90 [46]. Sequences per gene family were counted, normalized for length and alignment quality, and linked to pathways using MetaCyc [47]. Data were analyzed based on reads per kilobase of gene length (RPK) and reads per kilobase of gene length per million reads, where the total number of sequences in a sample = library size. HUMAnN2 pathway abundances were generated in terms of RPKs. HUMAnN2 processing of the metatranscriptomics samples was seeded with the bacterial genomes found from the corresponding metagenomics MetaPhlan2 results to optimize the correspondence between the 2 measures. CAZyme RPK abundances were generated using DIAMOND and the CAZyme (CAZy) database. Candidate pathways were used in subsequent analysis if the e value < 1e⁻²⁰, and RPK abundances were generated using the CAZyme gene length as found on the CAZy website: http://www.cazy.org/glycoside-hydrolases.html

Given a feature table of genes (DNA) or gene expression (RNA), we produced smoothed RNA and DNA values as well as relative expression values. “Smoothing” was estimated by substituting a small value in place of a 0 or missing value by “Laplace” (pseudocount) scaling, where the pseudocount is the sample-specific minimum nonzero value. To account for differences in microbial biomass that may be associated with diet [48,49], RNA/DNA relative expression was calculated with the HUMAnN2 rna_dna_norm.py script and is the smoothed value of RNA divided by the smoothed value of DNA and adjusted for differences in sequencing depth [50]. After the script adds pseudocounts for 0 entries, it computes the RNA/DNA relative expression per sample per feature for data generated for genes or transcripts identified in MetaCyc (https://metacyc.org/) or CAZyme database. The equation is as follows: REL(s,f), where s is the sample indicator and f is the feature.

REL(s,f) = RNA_(s,f) ∗ DNA_SUM_(s)/(RNA_SUM_(s) ∗ DNA_(s,f))

where

RNA_(s,f) is the RNA RPK value for sample, s, and feature, f

DNA_(s,f) is the DNA RPK value for sample, s, and feature, f

RNA_SUM_(s) is the sum of RNA_(s,f) over all features, f

DNA_SUM_(s) is the sum of DNA_(s,f) over all features, f

A filtered gene or transcript table was created by keeping genes (UniRef90 gene families) that had mean relative abundance >0.01%. Similarly, a filtered pathway table was created by keeping pathways that had mean relative abundance >0.03% (over pathways in the metagenomics and metatranscriptomics samples, respectively). For each participant, data matrices of the abundance of genes or transcripts or RNA/DNA ratios, gene or transcript families, and genes or transcripts in metabolic pathways were evaluated and exported for statistical analysis.

Statistical analysis

Although we recruited on the basis of BMI, this measure tends to be a poor indicator of body fat. Thus, analyses were conducted based on adiposity classifications as determined by DXA: high body fat mass for females ≥32.0% and males ≥25.0%. This resulted in recategorization of some participants with n = 29 comprising the normal/low fat mass group and n = 51 comprising the overweight/high fat mass group. The diet order together with sex, age, adiposity (defined as 0 if <25% for males or <32% for females, 1 if ≥25% for males or ≥32% for females) and calorie intake level were adjusted in the models below.

16S rRNA gene

Bacterial counts were filtered to 80% prevalence across participants, and then center log ratio was transformed [51]. Data were assessed using a linear mixed model (LMM).

Alpha and beta diversity

For 16S rRNA gene sequencing, alpha and beta diversity were calculated as previously reported [52]. Beta diversity of CAZymes was calculated using the Euclidean distance metric on filtered data (see below).

Data filtering at population level

For species analysis, metagenomic data had 100 species after filtering at >80% prevalence and 32 species at <80% prevalence. Metagenomic (DNA) and metatranscriptomic (RNA) measurements were normalized using a samplewise log-ratio centering method, and then RNA/DNA ratio was taken. At the population level, all variables were filtered at 80% prevalence. For RNA/DNA relative expression of bacterial metabolic pathways, we used a relative expression filtering criterion of >80% prevalence, removed all 55 superpathways, and put back 5 diet-related superpathways. The CAZyme RNA/DNA ratio data retained 192 variables after we filtered by RNA/DNA ratio >1 of variables identified in 80% of subjects. The distribution of the features was checked, and the log₂ transformation was applied where the data were not normally distributed.

Spearman rank correlation between RNA/DNA in LGL and HGL diets

We calculated the Spearman rank correlation between DNA and RNA metabolic pathway expression for the LGL and HGL diets, independently. Metabolic pathway expression was filtered, and a variable was removed if the prevalence of low-expressed pathways (RNA/DNA ratio < 1) had a prevalence of >80%. Spearman rank correlation was calculated on the DNA compared with RNA pathways demonstrated by the functional pathways.

Differential expression for metabolic pathways and CAZymes

LMMs were conducted for each feature (pathway and CAZymes), with the 2 time-point measurements. The diet order, together with sex, age, BMI (categorized as 0 if <25% for males or <32% for females, 1 if ≥25% for males or ≥32% for females) and calorie intake were adjusted in the models. Participant ID was used as a random effect. Features were deemed to be significantly different between diets if their mean difference between the diets exhibited a Benjamini–Hochberg-adjusted [53] P value (denoted here by P_adj) of <0.1.

Differential abundance of species

Species were identified using metagenomic sequencing and bioinformatic analysis using MetaPhlan [54] and as described above. LMMs were then conducted for each feature (species) with 2 time-point measurements and using the same approach as in the paragraph immediately above.

Association of HOMA-IR with microbial pathway expressed in either the HGL diet or the LGL diet

We used an LMM with HOMA-IR as the outcome associated with an interaction term to assess the pathways expressed in each diet. The equations were adjusted for the same variables and multiple testing as in the differential expression section above. “log₂(HOMA-IR) ∼ pathway + diet + diet ∗ path + age + sex + bodyfat + kcal + diet.seq + (1|ID),” where pathway is log₂(RNA/DNA) values.

Data and code availability

Analysis code presented in this study is available through the GitHub repository https://github.com/yzhanggmail/Bactocarb_Metanomics. This archive contains relevant analysis code as submodules, which include the code for mixed-effect model testing, Spearman rank correlation as well as scripts to generate the tables and figures are presented with detailed versions of other packages used. Metagenomics, metatranscriptomics, and 16S rRNA gene sequencing data are available at BioProjectID PRJNA1132291. Data elements from the CARB study are available for use for public access upon request to Dr. Marian Neuhouser, FHCC, Seattle, WA, United States, and execution of a data use agreement with the requestor’s institution.

Results

Characteristics and demographics of study participants and prescribed dietary intakes

Eighty participants provided stool samples at the end of each controlled feeding period (Table 1); 69 participants (35 men and 34 women) had a complete set of measures for 16S rRNA gene microbiome analysis and 64 had complete paired measurements for metagenomics and metatranscriptomics (Supplemental Figure 1). Mean prescribed dietary intakes for the 69 participants on the 2 diets are presented in Supplemental Table 1.

Biomarker analyses (HOMA-IR and CRP)

For the participants with microbiome data, CRP (mg/L; mean [SD]) on the HGL and LGL diets was 1.6 [2.7] and 1.3 [1.9], respectively. HOMA-IR on the HGL and LGL diets was 2.4 (1.8) and 2.5 (2.2), respectively. T tests on the log₂ scale comparing diet difference: P = 0.62 for HOMA ∼ diet and P = 0.83 for CRP across diets showed that neither measures were significantly different between diets.

Recovery of stool DNA, RNA, 16S rRNA gene sequences, and metagenomic and metatranscriptomic sequences

The mean ± SD yield of stool DNA was 174 mg/mL ± 122 (A₂₆₀/A2₈₀ = 2.0 ± 0.1; RIN = 5.6 ± 1.4), and after rRNA reduction, the yield was 175 mg/mL ± 148 (A₂₆₀/A2₈₀ = 1.9 ± 0.1; RIN = 3.5 ± 0.7; Supplemental Figure 2). For 16S rRNA gene sequencing (n = 69), we typically recovered a mean ± SD of 2.3 ∗ 10⁵ ± 9.1 ∗ 10⁴ sequences per sample after QC and the mean sequence length was 290 bp in samples taken at baseline and the end of the 2 dietary treatments. In the 69 participants with complete sample sets, we identified 9 bacterial phyla, 1 Euryarchaeota, and 161 genera. Samples for metagenomic and metatranscriptomic analysis (n = 64) were assessed at the end of each dietary treatment. Metagenomic and metatranscriptomic sequencing of samples resulted in a mean ± SD of 11.2 M ± 1.8 M and 54 M ± 15 M sequences per sample before QC and 11.1 M ± 1.8 M and 52 M ± 15 M after QC (Supplemental Table 2). The mean ± SD sequence length was 127 ± 28 bp for metagenomic and 122 ± 29 bp for metatranscriptomic data (Supplemental Table 2). We explored the impact of filtering the data (RNA/DNA ratio and RNA/DNA > or <1) on the distribution of normalized expression in the microbiome (>80% of participants) (Supplemental Tables 3–5). Among 215 pathways presented as relative % of RNA/DNA, 201 pathways were >80% prevalent (14 pathways were filtered and removed) (Supplemental Table 3). After pruning ∼90% of the superpathways, we obtained 147 pathways used for statistical testing (Supplemental Table 3). Our primary analysis was conducted with RNA/DNA ratios of pathways present in >80% of the samples after exploration of the distribution of pathways identified using RNA and DNA (Supplemental Tables 4 and 5).

GL had little effect on alpha and beta diversity of the microbial community

Based on 16S rRNA gene sequencing, we did not observe a difference in alpha and beta diversity between the 2 diet patterns (Figure 1A, B). After adjusting for diet order, age, body fat, energy intake, sex, and baseline measure 16S rRNA, 13 genera were significantly different (P_adj <0.05) after controlling the false discovery rate using the Benjamini–Hochberg procedure (Supplemental Table 6); 7 genera were more abundant in the LGL diet and 6 genera were more abundant in the HGL diet. Most of the genera were Firmicutes from either the Ruminococcaceae or Lachnospiraceae families (Supplemental Table 6). Using metagenomic sequencing, we identified 12 species that differed significantly between diets (Table 2). Five species were increased in the LGL diet, and 7 species, including Bilophila sp., Odoribacter splanchnicus, and Dorea longicatena, were increased in the HGL diet.

FIGURE 1.

(A) Beta diversity of microbial community measured using 16S rRNA gene sequencing. No significant diet differences in community structure (MiRKAT, P = 0.09). (B) Shannon diversity index in the LGL and HGL diets. There was no significant difference in diversity (t-test, P > 0.05). HGL, high glycemic load; LGL, low glycemic load.

TABLE 2.

Comparison of enrichment of species in each controlled diet.

Species name	β1 (se)	95% CI	Unadjusted P value	BH adjusted P2
Lachnospiraceae_bacterium_5_1_63FAA	1.32 (0.21)	(0.899, 1.739)	5.37E-08	5.37E-06
Parabacteroides_distasonis	−1.21 (0.20)	(−1.594, −0.822)	2.02E-07	1.01E-05
Ruminococcus_torques	−0.86 (0.19)	(−1.228, −0.478)	3.02E-05	8.20E-04
Anaerostipes_hadrus	1.23 (0.27)	(0.658, 1.74)	3.28E-05	8.20E-04
Odoribacter_splanchnicus	−0.37 (0.08)	(−0.535, −0.199)	1.13E-04	2.26E-03
Roseburia_unclassified	1.60 (0.40)	(0.769, 2.35)	7.30E-04	1.22E-02
Bilophila_unclassified	−0.43 (0.12)	(−0.671, −0.186)	9.58E-04	1.37E-02
Eubacterium_eligens	1.32 (0.38)	(0.579, 2.1)	1.27E-03	1.58E-02
Dorea_longicatena	−0.51 (0.15)	(−0.818, −0.203)	1.89E-03	2.06E-02
Parabacteroides_merdae	−0.38 (0.11)	(−0.603, −0.149)	2.06E-03	2.06E-02
Coprococcus_comes	−0.35 (0.11)	(−0.563, −0.122)	2.86E-03	2.60E-02
Coprococcus_catus	0.40 (0.13)	(0.138, 0.65)	3.52E-03	2.94E-02

We identified species using metagenomic sequencing and a linear mixed-effect model on center log-ratio transformed species (N = 12).

HGL, high glycemic load; LGL, low glycemic load.

¹A linear mixed-effect model “species ∼ diet + age +sex + bodyfat + kcal + diet.seq + (1|ID),” where species is log₂(DNA) values. A negative regression coefficient, β, coefficient corresponds to species abundance that increased in the HGL diet compared with LGL; a positive coefficient corresponds to species abundance that increased in the LGL diet compared to HGL.

²Benjamini–Hochberg adjusted P cutoff = 0.05.

Microbial expression detected differences in the carbohydrate composition of the diet

Expression of the hexitol fermentation pathway and the hexitol superpathway was enriched in the HGL diet compared with LGL diet ([β = −1.15; SE = 0.24; 95% confidence interval [CI]: −1.63, −0.67]; P_adj = 0.002) and ([β = −0.93, SE = 0.26; 95% CI: −1.455, −0.418]; P_adj = 0.037) (Figure 2A, Table 3). The expressed enzymes in the hexitol pathway were distributed across Bacteroidetes, Firmicutes, and Actinobacteria (Figure 2B). Expression of the L-lysine biosynthesis pathway was enriched in the LGL diet ([β = 0.20, SE = 0.05; 95% CI: 0.09, 0.30]; P_adj = 0.03), and starch degradation was enriched ([β = 0.27, SE = 0.10; 95% CI: 0.083, 0.459]; P_adj = 0.09) (Figure 2C, D, Table 3). Genes expressed (RNA) in pathways and genes (DNA) in pathways were typically correlated in the HGL and LGL diets, although not all pathways showed an association, and some were even inversely correlated (Supplemental Figure 3, Supplemental Table 7). The composition of expressed carbohydrate-active enzymes (CAZymes) differed significantly between diets (Figure 3A, P < 0.001). Interestingly, the expression of only 3 CAZymes and 1 starch-binding module was significantly enriched in the HGL diet, glucosyl hydrolase 13_3 containing pullulanase, GH130 containing 4-O-β-D-mannosyl-D-glucose phosphorylase, GH5_11 containing endo-β-1,4-glucanase/cellulase, and CBM53, a starch-binding module (Figure 3B, Table 4). The number of CAZymes enriched in the LGL diet was ∼3 times greater than in the HGL diet (Figure 3B). Expression of enzymes involved in starch, chitin, pectin, and cellulose degradation as well as their binding proteins were enriched in the LGL diet and largely from Firmicutes (Figure 3B, Table 4).

FIGURE 2.

(A) Pathways (RNA/DNA ratios) that are differentially expressed between diets for microbial metabolic pathways for linear mixed-effect model “pathway log2(RNA/DNA) value) ∼ diet + age +sex + bodyfat + kcal + diet.seq” + 1/subject, diet (orange highly expressed in LGL; blue highly expressed in HGL). Phylogenetic distribution of phyla expressing enzymes in (B) hexitol fermentation, (C) L-lysine biosynthesis, and (D) starch degradation. HGL, high glycemic load; LGL, low glycemic load.

TABLE 3.

The top 11 pathways ranked by BH adjusted P values in models testing for a difference between diets.

Name	β1 (se)	95% CI	Unadjusted P value	BH adjusted P2
P461 PWY hexitol fermentation to lactate, formate, ethanol, and acetate	−1.15(0.24)	(−1.627, −0.674)	1.24E-05	1.82E-03
PWY 5097 L-lysine biosynthesis VI	0.20(0.05)	(0.093, 0.304)	3.95E-04	2.90E-02
HEXITOLDEGSUPER PWY superpathway of hexitol degradation bacteria	−0.93(0.26)	(−1.455, −0.418)	7.46E-04	3.66E-02
PWY 5100 pyruvate fermentation to acetate and acetate II	−0.24(0.07)	(−0.384, −0.086)	2.18E-03	7.34E-02
PWY 5104 L-isoleucine biosynthesis IV	−0.51(0.16)	(−0.829, -0.186)	2.67E-03	7.34E-02
PWY 6936 seleno-amino acid biosynthesis	−0.63(0.20)	(−1.027, −0.22)	3.00E-03	7.34E-02
PWY 7237 myo, chiro, and scillo inositol degradation	−0.58(0.20)	(−0.976, −0.187)	5.18E-03	9.07E-02
P185 PWY formaldehyde assimilation III dihydroxyacetone cycle	0.77(0.27)	(0.24, 1.304)	5.70E-03	9.07E-02
PWY 6609 adenine and adenosine salvage III	−0.24(0.08)	(−0.398, −0.073)	5.83E-03	9.07E-02
PWY 6737 starch degradation V	0.27(0.10)	(0.083, 0.459)	6.17E-03	9.07E-02
PWY0 1319 CDP diacylglycerol biosynthesis II	−0.14(0.05)	(−0.237, −0.04)	7.50E-03	1.00E-01

HGL, high glycemic load; LGL, low glycemic load.

¹Linear mixed-effect model “pathway ∼ diet + age +sex + bodyfat + kcal + diet.seq + (1|ID),” where pathway is log₂(RNA/DNA) values. A negative regression coefficient corresponds to a pathway enriched in the HGL diet; a positive coefficient corresponds to a pathway enriched in the LGL diet.

²Benjamini–Hochberg adjusted P cutoff = 0.1.

FIGURE 3.

(A) Expression of CAZymes was significantly different between treatments (MiRKAT, P < 0.015; orange = LGL; blue = HGL), and (B) Volcano plot of linear mixed-effect model “CAZyme log₂(RNA/DNA) value) ∼ diet + age +sex + bodyfat + kcal + diet.seq” + 1/subject (orange highly expressed in LGL; blue highly expressed in HGL). CAZyme, carbohydrate-active enzyme; HGL, high glycemic load; LGL, low glycemic load.

TABLE 4.

Carbohydrate active enzymes (CAZymes) that were significantly differentially expressed between diets.

CAZyme Group	β1 (se)	95% CI	Unadjusted P value	BH adjusted P2	EC #	Enzyme
GH14	3.17(0.38)	(2.422, 3.911)	6.93E-12	1.33E-09	EC 3.2.1.2	β-Amylase
GH13_13	−2.70 (0.36)	(−3.406, −1.979)	2.92E-10	2.80E-08	EC 3.2.1.41	Pullulanase
GH152	3.15 (0.48)	(2.211, 4.093)	9.73E-09	6.22E-07	EC 3.2.1.39	β-1,3-Glucanase
PL1_1	2.59 (0.40)	(1.808, 3.378)	1.33E-08	6.37E-07	EC 4.2.2.2	Pectate lyase
GT47	1.68 (0.31)	(1.063, 2.299)	1.25E-06	4.79E-05	EC 2.4.1.-	xyloglucan β-Galactosyltransferase
GH19	2.40 (0.49)	(1.44, 3.379)	7.17E-06	2.29E-04	EC 3.2.1.14	chitinase
CBM49	1.16 (0.28)	(0.617, 1.701)	7.92E-05	2.17E-03		Crystalline cellulose binding
CBM18	1.70 (0.41)	(0.889, 2.508)	1.07E-04	2.56E-03		Chitin binding
GT34	1.40 (0.34)	(0.72, 2.076)	1.36E-04	2.91E-03	EC 2.4.1.-	UDP-Gal: galactomannan α-1,6-Galactosyltransferase
GH130	−0.30 (0.07)	(−0.442, −0.149)	1.69E-04	3.24E-03	EC 2.4.1.281	4-O-β-D-mannosyl-D-Glucose phosphorylase
CBM53	−1.11 (0.31)	(−1.719, −0.507)	6.21E-04	9.17E-03		Starch binding
GT75	1.32 (0.37)	(0.606, 2.053)	5.71E-04	9.17E-03	EC 2.4.1.	UDP-Glc: self-glucosylating β-glucosyltransferase
CBM 45	1.13 (0.31)	(0.516, 1.755)	6.05E-04	9.17E-03		Starch binding for GH14
GH13_6	1.11 (0.31)	(0.496, 1.733)	7.41E-04	1.02E-02	EC 3.2.1.98	glucan 1,4-α-Maltohexaosidase
GT106	1.23 (0.36)	(0.536, 1.939)	9.44E-04	1.21E-02	EC 2.4.1.351	UDP-β-L-rhamnose:rhamnogalacturonan I 4-α-Rhamnosyltransferase
GH5_11	−0.76 (0.25)	(−1.253, −0.262)	3.66E-03	4.40E-02	EC 3.2.1.4	endo-β-1,4-glucanase / cellulase

HGL, high glycemic load; LGL, low glycemic load.

¹A linear mixed-effect model “cazyme ∼ diet + age +sex + bodyfat + kcal + diet.seq + (1|ID),” where Cazyme is log2(RNA/DNA) values (after filtered by RNA/DNA > 1 of genes identified 80% of subjects). A negative regression coefficient (β) correspond to a pathway enriched in the HGL diet; a positive coefficient corresponds to pathways enriched in the LGL diet.

²Benjamini–Hochberg adjusted P cutoff = 0.1.

Microbial fatty acid metabolism is associated with a marker of insulin resistance in a refined-grain diet

In a previous analysis [13], we observed no differential effect on HOMA-IR by LGL diet compared with HGL diet. However, in this study, we tested whether the association of HOMA-IR with the expression of microbial functional pathways was effected by diet. We found a significant interaction between diet and the expression of 2 microbial functional pathways in their association with HOMA-IR. Specifically, the interaction of diet with CoA biosynthesis and diet with pantothenate (B₅) and CoA biosynthesis were both significant (P_adj = 0.035) and suggests that as the expression of these microbial pathways increase, HOMA-IR decreases in the LGL diet (β = −0.23; SE = 0.09; 95% CI: −0.40, −0.06) but increases on the HGL diet (β = 0.20; SE = 0.09; 95% CI: 0.02, 0.39) (Table 5). In contrast, after similarly modeling the association of CRP with microbial pathways and diet, we observed no significant associations.

TABLE 5.

The association of HOMA IR and expressed microbial pathways in each diet.

Name	β(se) for the effect of diet (LGL diet∗ pathway)1	95% CI LGL	β(se) for the effect of diet (HGL diet∗ pathway)1	95% CI HGL	Unadjusted P (diet∗pathway)1	BH adjusted P (diet∗ pathway)1
CoenzymeA biosynthesis I	−0.23(0.09)	(−0.404, −0.057)	0.20 (0.09)	(0.02, 0.387)	0.0005	0.035
Pantothenateand coenzyme A biosynthesis I	−0.31(0.11)	(−0.522, −0.09)	0.23(0.11)	(0.005, 0.451)	0.0003	0.035

The 2 pathways show significant differences when comparing the coefficient of pathway under LGL diet compared with the coefficient of pathway under HGL diet. With the LGL diet, the 2 pathways have a negative association with HOMA IR, whereas with the HGL diet, there is a positive association with HOMA IR. HGL, high glycemic load; LGL, low glycemic load.

¹Model “log₂(HOMA IR) ∼ β₁∗pathway + β₂∗diet + β₃∗diet∗path + β₄∗age +…,” output column “β (LGL diet ∗ pathway)” is “β₁+β₃” and column “β (HGL diet ∗ pathway)” is “β_1."

Discussion

The composition of the gut microbiome in healthy adults is generally stable, resilient [[55], [56], [57]], and driven by long-term dietary patterns [18,24,27,29,58]. Our 4-wk controlled dietary intervention revealed no significant difference in either species diversity (alpha diversity) or beta diversity when participants consumed the LGL and HGL diets. The LGL diet provided a mean intake of 55 g/d of dietary fiber from a diverse set of minimally processed foods, whereas the HGL diet provided a mean intake of 28 g/d. Previous studies have found that consuming complex carbohydrates (e.g., dietary fibers, resistant starch) often increases the diversity and prevalence of particular bacterial taxa [20,[31], [32], [33], [34],[59], [60], [61], [62], [63], [64], [65], [66], [67]]. However, a recent meta-analysis of dietary interventions found mixed results on how fiber affects gut microbiome diversity and composition [68]. For example, a randomized clinical trial found no significant difference in either alpha diversity or glycemic control when participants were fed bread with a low GI for 7 consecutive days [69]. We and others have also observed no change in alpha diversity in response to other short-term plant-based, phytochemical or fiber interventions [36,[70], [71], [72]].

Diets rich in dietary fiber have been associated with a lower risk of chronic inflammation-related health outcomes [73]. Five species that ferment fiber to short-chain fatty acids were in higher abundance in the LGL diet compared with the HGL diet. For example, Eubacterium eligens and 2 species of Roseburia found in the Lachnospiraceae family ferment fiber to butyrate. E. eligens has been positively associated with dietary fiber intake [74] and produces butyrate, often associated with lower inflammation. In prospective cohort studies, a lower abundance of E. eligens in feces was associated with increased risk of colorectal cancer (CRC) [[75], [76], [77]] supporting the link to butyrate and CRC [78,79]. Butyrate-producing Roseburia sp. have also been associated with reduced chronic inflammation and positive health outcomes [80,81].

In contrast, many of the genera that were increased in the HGL diet have been associated with higher systemic inflammation and obesity [82]. Bilophila sp. are also positively associated with animal-based diets enriched in fats and different aspects of bile acid metabolism. In our study, although the percent energy from carbohydrate, protein, and fat was held constant across the 2 diets, there was greater animal protein intake in the HGL diet (18%) to compensate for the increased protein in vegetables and whole grains on the LGL diet (47%). Bilophila sp. metabolize the taurine in taurine-conjugated bile acids to hydrogen sulfide [83]. O. splanchnicus, enriched in the HGL diet, also produces hydrogen sulfide [84]. Both secondary bile acids and hydrogen sulfide are genotoxic and are associated with inflammation, intestinal barrier dysfunction, and glucose dysregulation [85].

On a typical Western diet, 2–10 g/d of unabsorbed sugars and sugar alcohols reach the colon [[86], [87], [88], [89], [90]] and are substrates for microbial metabolism. Sugar alcohols are polyols enriched in refined-grain diets. In our study, polyol metabolism (microbial hexitol fermentation pathway) was higher in the HGL diet and may reflect dietary components that were not digested and absorbed in the small intestine that reached the colon [91]. Other studies have highlighted the relationship between polyols and the gut microbiome [[92], [93], [94], [95]] and showed an altered abundance of the hexitol fermentation pathway associated with a more processed diet [72,96,97]. One microbial fermentation product from the hexitol pathway, ethanol, has been associated with increased systemic inflammation [98]. In contrast, the LGL diet enriched the microbial pathway that produces L-lysine from aspartate. L-lysine is an essential amino acid in humans; it must be ingested or derived from gut bacteria that synthesize L-lysine [[99], [100], [101]]. In a previous metabolomic assessment of this dietary intervention, we found that levels of aspartate were lower in the LGL diet suggesting that microbial metabolism may impact systemic levels of certain circulating metabolites [15].

A diet enriched in refined grains with lower structural complexity may ultimately lead to reduced functional complexity in the microbiome often associated with Westernized, low-fiber diets [58,[102], [103], [104]]. We showed that the expression of CAZymes reflected the variety and complexity of dietary carbohydrate intake. In this study, mean daily fiber intake in the HGL diet was lower than in the LGL diet but higher than that of typical United States diets [105]. Nonetheless, the lower structural complexity in processed foods was reflected in the expression of fewer types of CAZymes in the HGL diet. More processed foods often contain additives, such as thickeners, emulsifiers, or sugar alcohols [106]. When participants consumed the HGL diet, dietary additive metabolism dominated the expression profiles enriched in the glucosyl hydrolase family (GH13-3), which includes pullulanase [107]. Pullulan is used as a binder, thickener, and coating agent in foods and as a prebiotic [[108], [109], [110]]. Although these food additives are naturally occurring compounds, they replace the array of diverse monosaccharides and glycosidic bonds found in whole foods [111]. In contrast, in the LGL diet, microbial expression of a variety of glucosyl hydrolases and carbohydrate-binding modules reflected varied plant sources with a diversity of oligosaccharides linkages. The CAZymes expressed identified linkages found in complex starches, lamarin, pectin, and chitin, all part of the LGL diet. Others have shown dramatic effects of food additives on the microbiome [[112], [113], [114], [115], [116]]. A study in humanized mice showed that a refined carbohydrate diet resulted in an irreversible loss in glucosyl hydrolase diversity after multigenerational exposure to refined diets [102]. Cross-sectional studies of acculturation to a United States diet showed lower diversity in the gut microbiome [117,118]. The lower diversity of glycan-degrading enzymes we found in the HGL diet is consistent with lower structural complexity of processed foods found in a typical United States diet [103,119].

Diets high in refined grains and simple carbohydrates are typically associated with higher HOMA-IR, a marker of insulin resistance [[120], [121], [122]]. In previous analyses, we found that the HGL diet resulted in higher postprandial glucose and insulin responses compared with the LGL diet but not in higher HOMA-IR [13]. Here, we show that even in a relatively healthy, nondiabetic study population, microbial metabolism of different diet types varies with markers of host insulin resistance. We found that microbial biosynthesis of pantothenic acid (vitamin B₅) was positively associated with HOMA-IR in the HGL diet. Humans are auxotrophic for pantothenic acid and rely on sources from the diet or the gut microbiome [123]. Pantothenic acid is required in humans for the synthesis of coenzyme A and acyl-carrier protein [124] and plays a role in host fatty acid synthesis and processing of large organic molecules, such as carbohydrates, lipids, and proteins. Microbial ethanol biosynthesis, as part of the hexitol fermentation pathway, was also enriched in the HGL diet. Others have found a positive association with the microbial ethanol biosynthesis pathway and HOMA-IR [125]. Because we show that microbial metabolism was manipulable by diet, the identification of microbial metabolic pathways linked to the development of insulin resistance may be a useful strategy to lessen insulin resistance and related disorders [126].

Our study had both strengths and weaknesses. We used a gene-normalized expression (metatranscriptomic) approach to focus on microbial expression rather than gene content alone, thereby capturing microbial responsiveness to dietary intake [127]. A strength of this approach is that it accounts for differences in microbial gene copy numbers that may vary between individuals. This facilitated a more robust comparison of microbial gene expression and response to the intervention. The randomized, controlled intervention with all individuals consuming the same food for the 2 dietary patterns is also a study strength. The crossover study design, with each person serving as their comparison, further reduced confounding by nondietary factors. This design is especially relevant to measures of the microbiome, considering the high interindividual variation in the composition of the microbiome can obscure response in parallel arm interventions. We did not directly measure the impact of the 28-d washout period on the microbial community; however, prior studies have shown this to be adequate for the microbiome to return to baseline [36,[128], [129], [130]]. Another strength was that the eucaloric dietary intake was monitored and adjusted to minimize weight change. Both male and female participants maintained their usual weights during the study [11], allowing us to focus on the impact of diet on the microbiome, without weight change as a confounder. To maintain similar carbohydrate intake between the 2 diets, we increased the fructose content of the LGL diet [11]. Although we saw an impact of increased fructose on human lipid profiles [14], we did not see effects on microbial fructose metabolism. Recently, others have shown that microbiome acetate production impacts host de novo lipogenesis when the rate of fructose ingestion exceeds small intestine uptake capacity and fructose reaches the colon [131]. Although we did not measure increased expression in acetate pathways in the colon, we cannot completely rule out that microbially produced acetate may have contributed to the altered host lipid profiles. Although some may view a higher intake of fructose as a lower-quality sweetener, and a weakness in the low glycemic diet quality, our results support the observation that fructose intake within a high-fiber diet minimizes the effects of excessive fructose consumption and highlights the benefits of using a whole-food diet intervention as opposed to one focused on dietary components [132].

Evaluation of normalized gene expression in the microbial community could be used as a biomarker of dietary quality relevant to an individual’s health. However, many studies to date have focused on adding only a single fiber source instead of altering carbohydrate quality within the context of a whole-food diet [[31], [32], [33], [34]]. In our study, microbial gene expression reflected both potential food additives in the HGL and the diversity of whole grains in the LGL diet. Assessment in a long-term setting may identify which aspects of microbial metabolism impact host physiology to alter health outcomes in a targeted and personalized manner.

Author contributions

The authors’ responsibilities were as follows – JWL, MLN: designed the parent research study and wrote the paper; MAH: was responsible for microbiome-related analysis including fecal sample preservation from the parent study, microbial laboratory oversight and design, and bioinformatic pipelines for analysis of 16S rRNA gene, metagenomics, and metatranscriptomics; QC: designed statistical analysis; OK: performed laboratory analysis; MAH: wrote paper and MAH and JWL had primary responsibility for final content. TWR, YZ, SN: analyzed data and performed statistical analysis; LL: provided data cleaning and manuscript preparation; KRC: provided programming and bioinformatic pipeline building and analysis; DR, AS, SN, YZ, TWR, KRC, MK: critically reviewed the manuscript; and all authors read and approved the final manuscript.

Data availability

Data described in the manuscript, code book, and analytic code will be made publicly and freely available without restriction in the Sequence Read Archive (http://www.ncbi.nlm.nih.gov/sra/.). Data elements from the CARB study are available for use public access upon request to Dr. Marian Neuhouser, FHCC, Seattle, WA, United States, and execution of a data use agreement with the requestor’s institution.

Funding

This study was supported by NIH grants U54 CA116847, R01 CA192222, and P30 CA015704.

Conflict of interest

The authors report no conflicts of interest.

Appendix A. Supplementary data

The following is the Supplementary data to this article: