A population-scale analysis of 36 gut microbiome studies reveals universal species signatures for common diseases

The gut microbiota is increasingly recognized as a key regulator of host health. Although the gut community is generally resilient to acute perturbations, chronic exposure to pollutants, stressors, and diseases can induce dysbiosis—alterations in composition that may favor more virulent microorganisms and harm the host. With advances in amplicon and whole-metagenome sequencing, dysbiosis has been reported across many common conditions, including autoimmune disorders, cardiometabolic diseases, infections, psychiatric disorders, and cancers. However, across-study comparisons remain difficult due to heterogeneous reference databases, variable taxonomic annotation accuracy, and diverse experimental and analytical workflows. The precise contribution of microbial dysfunction to disease also remains incompletely understood. This study aims to systematically characterize gut microbiota alterations across multiple diseases using publicly available datasets in a unified framework, to identify overall patterns of disease-associated microbiota shifts via large-scale meta-analysis.

Prior work has documented widespread associations between gut microbiome composition and a broad range of diseases, and several meta-analyses have suggested both disease-specific and shared microbial responses across conditions. Nevertheless, inconsistent methodologies, batch effects, and non-uniform reference databases have limited the comparability of findings. Earlier studies highlighted reductions in short-chain fatty acid (SCFA)–producing taxa (e.g., Faecalibacterium, Roseburia) as common features of dysbiosis, and enrichment of opportunistic pathogens (e.g., Fusobacterium, Enterococcus, Escherichia) in multiple diseases. Meta-analytic and machine-learning approaches have been proposed to enhance reproducibility and extract generalizable microbial signatures for diagnosis across diseases under large-scale designs. Building on this literature, the present work applies a unified reprocessing and statistical pipeline to thousands of metagenomes to resolve universal and disease-shared species-level signatures.

Study identification and data collection: Published gut metagenome case-control studies were searched in PubMed and Google Scholar with exhaustive keywords (as of January 2022). Inclusion criteria were: (1) case-control design of disease, (2) samples from Chinese individuals, and (3) availability of raw metagenomic data. Of 45 eligible studies, 8 were excluded (<50 total samples or case/control not from same batch), leaving 36 studies with 6,314 fecal metagenomes. Metadata were obtained from original articles, public repositories (NCBI SRA/ENA/CNGB), or by contacting authors. Participants were excluded for non-standard disease definitions, absence of baseline in longitudinal designs, or BMI <17 or >30 kg/m² with comorbidities. New autoimmune cohort: An independent cohort (ethics-approved, informed consent) included 95 RA, 73 SLE, 65 pSS patients, and 118 healthy controls recruited from two hospitals. Exclusions included prior IBD, diabetes, severe hypertension/obesity/metabolic syndrome, renal cancers, abnormal liver/kidney function, and recent antibiotics/probiotics (4 weeks). Fecal samples were collected, stored at −80°C, DNA extracted (TIANGEN kit), library prepped (NEBNext Ultra), QC’d (Agilent 2100), sequenced on Illumina NovaSeq (150 bp PE). Quality control and human read removal matched the public dataset pipeline. Read processing and taxonomic profiling: Raw reads were quality-filtered with fastp (trim low-quality tails <Q20 or N; remove reads <145 bp). Human reads were removed by mapping to GRCh38 with Bowtie2; samples with insufficient high-quality reads were excluded. Taxonomic profiles were generated using MetaPhlAn 4, which bins species into species-level genomic bins (SGb) and references ~22,000 known species and 5,000 strains. Within-study analyses: For each case-control comparison (40 comparisons across disease subtypes/severities), species richness (observed species) and Shannon diversity were compared using rank-sum tests, adjusting for age, sex, and BMI when available. Community structure differences were tested via PERMANOVA (Bray–Curtis; 1,000 permutations). Random forest models (R package randomForest; 100 trees) were trained within each study to classify cases vs controls; AUCs were computed via ROC analysis (pROC). Meta-analysis and confounder-adjusted association testing: To identify disease-associated taxa across studies, a random-effects meta-analysis (metafor) was combined with MaAsLin 2. Relative abundances were arcsine–square-root transformed; effect sizes were estimated using Hedge’s g under a random-effects model. Linear mixed-effects models (lme4) included study as a random effect. Multiple testing was controlled by Benjamini–Hochberg. Taxa were considered differentially abundant if both (1) random-effects meta-analysis p<0.05 and (2) MaAsLin 2 p<0.05 after adjusting for age, sex, and BMI (where available). Multivariate analyses: Community clustering used Jensen–Shannon divergence with PAM; the average silhouette width (ASW) evaluated cluster support (ASW<0.3 indicating poor support). Ordination used PCoA and distance-based RDA (Bray–Curtis). PERMANOVA effect sizes (R²) were computed with vegan::adonis. Correlations used Spearman’s rho. Prediction and health index: Cross-validated random forests were trained on the 277 disease-associated species to classify cases vs controls and high-risk vs controls. LASSO models were also evaluated. A random forest–based gut microbiome health index (GMHI) for each sample aggregated predictions across ten RF classifiers. External validation used independent public cohorts (bipolar depression, CRC, ESRD) and the new autoimmune cohort (RA, SLE, pSS).

- Across 40 case-control comparisons (subtypes/severities separated), many diseases showed significant decreases in species richness and Shannon diversity in cases vs controls. Twelve comparisons had lower diversity in cases; only two showed higher diversity. Crohn’s disease showed >10% decreases in both richness and diversity across two comparisons. Several conditions (COVID-19, pulmonary tuberculosis, hypertension, SLE, liver cirrhosis, gout, Graves’ disease, ankylosing spondylitis) exhibited >10–20% decreases in richness/diversity, while Parkinson’s disease showed increased richness/diversity in some cohorts. - PERMANOVA indicated significant disease-associated compositional shifts in 27/40 comparisons (p<0.05). Within-disease random forests achieved AUC>0.7 in 28/40 comparisons, with an average AUC of 0.759. - Aggregated PERMANOVA across all samples showed disease status significantly but modestly explained gut microbiome variation (R² = 0.43%, p<0.001), while sex+age+BMI together explained 0.26% (p<0.001), supporting shared disease-related signatures. - Meta-analysis identified 277 disease-associated species: 194 enriched in healthy controls and 83 enriched in diseased subjects. Control-enriched taxa were dominated by Firmicutes and included SCFA producers such as Roseburia, Ruminococcus, and Faecalibacterium, as well as Bacteroides and Alistipes species. Disease-enriched species included opportunistic pathogens such as Streptococcus spp., Enterococcus spp., Escherichia coli, Fusobacterium spp. (e.g., F. nucleatum, F. varium), Hungatella hathewayi, Flavonifractor plautii, and several Lactobacillus species. - At the genus level, 107 genera differed between diseased and healthy individuals (73 control-enriched, 34 patient-enriched). SCFA-associated genera (Faecalibacterium, Roseburia, Butyricimonas) were depleted in disease, while harmful genera (Fusobacterium, Escherichia, Enterococcus) were enriched. - A random forest using the 277 species achieved AUC 0.776 (95% CI, 0.764–0.787) for cases vs controls and AUC 0.825 (95% CI, 0.806–0.845) for high-risk patients vs controls. Key discriminative species included Ligilactobacillus salivarius, Ruminococcus gnavus, Clostridium symbiosum, Streptococcus pneumoniae, Eggerthella lentus, Fusobacterium mortiferum, Blautia producta, Blautia hansenii, and Peptostreptococcus stomatis. LASSO performed worse (species-level AUC 0.735; genus-level AUC 0.705). - GMHI differentiated health states: mean±SD 0.655±0.151 (diseased) vs 0.491±0.141 (healthy) (Mann–Whitney P < 2.2×10⁻¹⁶). Of samples with GMHI ≤0.4, 81.0% were healthy (727/898); with GMHI >0.6, 80.6% were nonhealthy (236/292). GMHI correlated with richness (ρ=0.37) and diversity (ρ=−0.36). - External validation: Using the original RF classifier, AUCs were 0.637 (95% CI, 0.533–0.741) for bipolar depression, 0.838 (0.736–0.940) for colorectal cancer, and 0.836 (0.786–0.887) for end-stage renal disease. In an autoimmune cohort, AUCs were 0.555 (0.459–0.651) for RA, 0.638 (0.545–0.730) for SLE, and 0.717 (0.610–0.823) for pSS.

This population-scale reanalysis using a unified pipeline demonstrates that many common diseases are associated with decreased gut microbial richness and diversity and significant alterations in overall community structure. The identification of 277 universal disease-associated species supports the concept of shared microbial signatures across diseases—characterized by depletion of SCFA-producing commensals and enrichment of opportunistic pathogens. Predictive modeling using these signatures achieved robust cross-disease classification and generalized, with varying performance, to independent cohorts, performing best for high-risk diseases such as CRC and ESRD and less well for psychiatric conditions and RA where effect sizes are smaller. The work underscores the value of standardized processing and meta-analytic frameworks for deriving reproducible microbiome-disease associations and provides a resource for future investigations into mechanisms and potential microbiome-informed health assessments.

By reprocessing 6,314 fecal metagenomes from 36 case-control studies through a unified meta-analysis pipeline, this study delineates universal, cross-disease gut microbial signatures at species and genus levels. The consistent depletion of SCFA producers and enrichment of potential pathogens across diseases, together with robust classifier performance and a microbiome health index, highlight the translational potential of these signatures for risk stratification and personalized disease management. The findings emphasize the importance of data sharing and standardized analytic methods to achieve broadly applicable insights. Future research should extend validation across diverse populations, refine causal inference (e.g., longitudinal and interventional designs), and elucidate mechanistic pathways linking key taxa to host physiology and disease.

- Cohorts were restricted to Chinese populations, which may limit generalizability to other ancestries and environmental contexts. - Heterogeneity in study designs, sampling, sequencing, and available metadata across included studies persists despite unified reprocessing, and some analyses could adjust for confounders (age, sex, BMI) only when metadata were available. - Disease status explained a modest fraction of overall microbiome variance (R²≈0.43% across all samples), indicating small effect sizes at the population level. - Predictive performance varied by disease, with lower AUCs for psychiatric disorders and RA, suggesting weaker or more heterogeneous microbiome signals in these conditions. - The case-control, largely cross-sectional nature of the data limits causal inference regarding microbiome changes and disease onset or progression. - Some taxonomic and methodological inconsistencies in public datasets (e.g., differing databases/annotation accuracy) remain potential sources of bias, though addressed by MetaPhlAn 4 and meta-analysis.