A randomized controlled trial for response of microbiome network to exercise and diet intervention in patients with nonalcoholic fatty liver disease

Nonalcoholic fatty liver disease (NAFLD) is highly prevalent worldwide and is closely associated with type 2 diabetes through shared mechanisms such as insulin resistance. Modulating the gut microbiota is a promising target for NAFLD treatment because dysbiosis has been linked to NAFLD and T2D through inflammation, intestinal barrier disruption, and bacterial translocation. While low-carbohydrate diets and aerobic exercise can reduce hepatic fat and may alter gut microbial composition and function, mechanisms remain unclear and inter-individual variability in response is substantial. Traditional analyses may miss community-level interactions and keystone taxa that drive microbiome structure and function. This study investigates how aerobic exercise and/or a fiber-enriched low-carbohydrate diet affect gut microbiota composition, function, and co-occurrence networks in NAFLD patients with impaired glucose metabolism, and whether personalized baseline microbial networks can predict individual responses in liver fat reduction.

Prior studies have shown: (1) Dysbiosis of gut microbiota is associated with NAFLD and T2D, potentially via inflammatory pathways and barrier dysfunction. (2) Low-carbohydrate diets can rapidly shift gut microbiota and improve hepatic steatosis. (3) Aerobic exercise reduces hepatic fat and has been reported to increase microbial diversity and alter composition and function. (4) There is considerable heterogeneity in response to exercise and diet, with up to 50% showing low or no microbiota change after exercise; genotypic and phenotypic factors may contribute. (5) Microbial ecosystems are structured by keystone taxa and interaction networks; network properties such as connectivity and robustness are linked to ecosystem function and disease progression. These findings motivate analysis beyond abundance to include community networks and personalized microbial interactions.

Design: Randomized, single-blinded (researchers), controlled, four-arm trial of approximately 8.6 months (range 6.5–11.1). Arms: aerobic exercise (AEx), fiber-enriched low-carbohydrate diet (Diet), combined aerobic exercise plus diet (AED), and no intervention (NI). Recruitment: 115 participants from 7 health centers in Shanghai Yangpu district; 85 completed the trial; 76 provided paired stool samples for 16S rRNA sequencing; 42 paired samples underwent metagenomics. Trial registered (ISRCTN 42622771; retrospectively). Ethics approved; informed consent obtained. Participants: Men and women aged 50–65 years with impaired fasting glucose or impaired glucose tolerance; NAFLD diagnosed by 1H MRS (liver fat >5%); limited alcohol intake; exclusion included BMI >38 kg/m2, serious comorbidities, diabetes (type 1 or 2), mental illness; postmenopausal criteria for women. Interventions: AEx: supervised progressive aerobic exercise (Nordic brisk walking, stretching, group exercises) 2–3 sessions/week, 30–60 min/session at 60–75% VO2max, with warm-up/cool-down. Diet: daily prepared lunch (~30–40% daily energy): 37–40% carbohydrate with 9–13 g fiber plus 5 g soluble fiber, 35–37% fat (SAFA ~10%, MUFA 15–20%, PUFA ~10%), 25–27% protein; other meals self-prepared per nutritionist guidance; maintain physical activity. AED: both AEx and Diet as above. NI: maintain usual lifestyle. Outcomes and measurements: Primary outcomes previously reported were hepatic fat content (HFC) by proton MRS and glucose metabolism (HbA1c). Current report: third primary outcome, gut microbiota composition. Additional measures: body fat mass (DXA), serum biomarkers, fecal short-chain fatty acids (SCFAs) by GC. Microbiome assays: 16S rRNA gene sequencing (V3–V4) with QIIME2/DADA2; ASVs defined; taxonomy by SILVA v132; rarefaction to 10,000 reads/sample. Absolute bacterial abundance assessed by qPCR of 16S rRNA gene. Shotgun metagenomics on a subset (42 pairs) with assembly and annotation; functional profiling to KEGG pathways. Network analyses: Co-abundance groups (CAGs) constructed from 279 ASVs present in ≥20% samples using SparCC correlations (bootstrap n=100; |r|>0.4, p<0.05), leading to 35 CAGs. Co-occurrence networks (Spearman, p<0.05) built at baseline and follow-up for each arm to assess connectivity (edge number, mean degree) and robustness (node removal simulations; robustness statistic R from fragility curves, repeated 1000 times). Scale-free property and hub (keystone) CAGs identified by degree distribution (>4% of total connections). Personalized network (Single SparCC) analysis: Developed a method combining SparCC with sample-specific networks (SSN) to infer individual baseline microbial interaction networks from metagenomics and 16S data. Network attributes (edge number, mean degree) used to predict responders (HFC decrease >5%) vs low/non-responders via (a) unsupervised ROC based on edge number, and (b) supervised models (LASSO/Elastic-Net) using edge matrices as features. Also tested baseline ASV abundances for prediction. Associations between network edges and HFC change assessed by linear regression. Statistics: ANCOVA for repeated measures for alpha diversity and absolute abundance controlling for weight change, baseline value, intervention duration, with Sidak correction; PERMANOVA for beta diversity; LEfSe for differential taxa and KEGG pathways (LDA>2, p<0.05); partial Spearman correlations between ASV changes and clinical parameters (adjusted for body weight and fat mass); significance at p<0.05; visualization in Cytoscape.

- Completion: 85 completed intervention; paired 16S data from 76; metagenomics from 42. - Hepatic fat content (from prior report, for context): HFC decreased by −24.4% (AEx), −23.2% (Diet), and −47.9% (AED) vs +20.9% increase (NI), p<0.001; proportion with HFC decrease: AED 91%, AEx 68%, Diet 86%; NI 72% increased. - Diversity and composition: Shannon alpha diversity significantly decreased in NI over time (NI 0 m vs 8.6 m p=0.031), while maintained in intervention groups; post-intervention Shannon higher vs NI: AED p=0.011, AEx p=0.007, Diet p=0.025. Weighted UniFrac PCoA showed significant differences between NI and each intervention after intervention (PERMANOVA p<0.05). Absolute bacterial load by qPCR did not differ within or between groups. - Differential taxa (LEfSe, adjusted p<0.05, log2FC>2): AED vs NI had 15 enriched and 5 decreased ASVs; AEx had 13 enriched; Diet had 9 enriched and 6 decreased. ASVs increased across all intervention groups included Bacteroides (ASV2077 p-values: AED 0.018; AEx <0.01; Diet 0.032) and (ASV2513: AED 0.012; AEx <0.01; Diet 0.013), and Ruminococcus (ASV3942: AED 0.030; AEx 0.024; Diet <0.01). Lachnospiraceae ASV5361 increased in AEx (p=0.027) and AED (p=0.037); Bacteroides ASV2440 increased in Diet and AED (both p<0.01). Some ASVs from the same family showed divergent changes. - Clinical correlations (partial Spearman adjusted for body weight and fat mass, n=46): ASV2468 (Bacteroides; r=−0.31, p=0.040), ASV3307 (Ruminococcaceae; r=−0.32, p=0.030), ASV4538 (Lachnospira; r=−0.37, p=0.012) correlated with reductions in HFC. ASV478 (Phascolarctobacterium; r=−0.32, p=0.033), ASV1715 (Alistipes; r=−0.35, p=0.022), ASV5195 and ASV5305 (Lachnoclostridium; r=−0.30, p=0.045; r=−0.34, p=0.023) correlated negatively with HbA1c; ASV776 (Erysipelotrichaceae) correlated positively with HbA1c (r=0.36, p=0.016). SCFA levels did not significantly change over time, but multiple ASVs correlated with SCFAs (e.g., Faecalibacterium ASV3718 negatively with 4/6 SCFAs; Bacteroides ASV1989 positively with 4 SCFAs). - Functional pathways (metagenomics; LEfSe LDA>2, p<0.05): 64 KEGG pathways differed between intervention groups and NI, mainly in carbohydrate, energy, glycan, lipid, amino acid, and cofactor/vitamin metabolism. AED exhibited more differential pathways than AEx or Diet. Most pathways were more abundant in NI; glycan biosynthesis and metabolism was more abundant in intervention groups. Within lipid metabolism, sphingolipid metabolism showed notable differences. - Co-occurrence networks (CAG-level): Post-intervention, edge numbers and mean degree increased vs baseline in AED, AEx, and Diet but not NI. Network robustness increased in intervention groups and decreased in NI; robustness score increase largest in AEx (+29.3%), followed by Diet (+12.8%) and AED (+3.1%). Degree distributions indicated scale-free networks with hubs; four hub CAGs (CAG6, CAG7, CAG8, CAG28) accounted for >4% of connections. Their abundances increased significantly post-intervention in AED vs NI; dominant ASVs within these hubs belonged to Bacteroides, Ruminococcaceae, Alistipes, Subdoligranulum. - Personalized prediction: Baseline ASV abundances modestly predicted responders only in Diet (AUC 0.65; specificity 0.58; sensitivity 0.67), not in AED (AUC 0.53) or AEx (AUC 0.52). Single SparCC personalized baseline networks showed responders tended to have more species interactions. Linear regression: edge number significantly predicted HFC change in AEx (metagenomics; R2=0.45, p=0.0166), with trends in AED and Diet. Unsupervised ROC using edge number achieved AUCs ~0.70 (AED 0.77; AEx 0.74; Diet 0.72). Supervised LASSO improved AUCs (AED 0.90; AEx 0.94; Diet 0.76). Using 16S data, similar patterns held; baseline edge numbers correlated with HFC change in AEx and AED but not Diet. Age, weight, BMI were generally poor predictors of responsiveness, except BMI in AED.

This randomized trial in NAFLD patients with impaired glucose metabolism demonstrates that lifestyle interventions modulate the gut microbiome at both compositional and systems (network) levels. Exercise and diet maintained alpha diversity relative to NI, where diversity declined alongside worsening liver fat, suggesting interventions counteract disease-associated microbiome deterioration. Taxonomic shifts included increases in Bacteroides and Ruminococcus ASVs across interventions, taxa implicated in energy extraction and SCFA production; several ASVs correlated with reductions in hepatic fat and HbA1c and with SCFAs, linking microbiota changes to metabolic improvements. Functional profiling showed broad pathway differences, with glycan biosynthesis/metabolism enriched in intervention groups and sphingolipid metabolism notably altered, indicating systemic metabolic reprogramming of the microbiome. Network analyses revealed enhanced connectivity and robustness after interventions, particularly with exercise, indicating a more stable and resilient microbial ecosystem. Scale-free properties and identification of hub (keystone) CAGs suggest that interventions strengthen interactions among key taxa rather than merely altering diversity. The divergence wherein diet better predicted response using ASV abundances, but exercise yielded the greatest increase in network robustness, supports a model where diet directly reshapes microbial composition via nutrients whereas exercise modulates interspecies interactions through host-mediated factors, improving ecosystem stability and host metabolism. Importantly, personalized baseline microbial interaction networks predicted individual response in hepatic fat reduction, especially for exercise, outperforming simple demographic or anthropometric predictors. This indicates that network-level features capture critical ecological interactions relevant to intervention efficacy and could inform personalized treatment strategies for NAFLD.

An 8.6-month aerobic exercise and/or fiber-enriched low-carbohydrate dietary intervention in NAFLD with prediabetes maintained microbiome diversity, altered key taxa and functions, and strengthened gut microbial co-occurrence networks. Combined intervention diversified and stabilized keystone taxa; exercise or diet alone increased network connectivity and robustness. Personalized baseline microbial networks predicted individual hepatic fat responses, especially to exercise, suggesting clinical utility for tailoring interventions. Future research should validate these findings in larger, independent cohorts, refine network-based predictive models, and elucidate mechanistic links between network stability, keystone taxa, and host metabolic outcomes.

- Metagenomics subset sample size was relatively small, limiting power and generalizability. - Heterogeneity across trials in diet/exercise contents and intensity complicates direct comparison with other studies. - Trial was retrospectively registered; predictive analyses using personalized networks were not planned a priori and lacked independent validation. - Personalized prediction requires sufficient reference samples to construct robust sample-specific networks.