Evidence-gated integration of microbiome, metabolome, and host transcriptome data reveals a gap between reproducible signals and cross-layer mechanistic evidence in hepatocellular and colorectal cancer

Introduction Microbiome multi-omics studies frequently identify microbial, metabolic, and host transcriptional associations, but their reproducibility and support for connected biological mechanisms remain uncertain. We evaluated these relationships in hepatocellular carcinoma (HCC) and colorectal cancer (CRC) using an eight-dimensional, non-compensatory evidence-gating framework. Methods Nine public HCC datasets spanning microbiome, metabolome, and host transcriptome profiles were analyzed alongside CRC microbial and metabolomic data and separate host transcriptomic datasets. Analyses included cross-cohort replication, participant-matched taxon–metabolite associations, host transcriptional responses, sensitivity testing, and independent evidence scoring. Complete mechanistic eligibility required support across all eight evidence dimensions. Results In HCC, 13 genera showed concordant directions and nominal significance across conventional leave-one-cohort-out analyses, but none met the adjusted external validation criterion. Six longitudinal plasma feature-window signals and 12 differential liver metabolites were identified without cross-specimen concordance under false discovery rate (FDR) correction. Three hepatic transcriptional programs replicated across independent bulk datasets, but none of the four HCC candidate families met all eight evidence requirements. In CRC, four bacterial species replicated in held-out disease-feature testing. Paired microbial and fecal metabolite profiles from 347 participants yielded 18 replicated taxon–metabolite associations, of which 10 met robustness criteria and eight retained support after joint-taxon adjustment. The Peptostreptococcus stomatis–glycocholate association replicated in held-out testing (adjusted residual r = 0.315, q = 0.00232) and remained significant after joint-taxon adjustment (β = 0.209, q = 0.0144) and unified-partition analysis with global Benjamini–Hochberg correction (held-out q = 0.0303). Reanalysis of glycocholic acid (GCA) treatment transcriptomic data identified increased SOX14, ZDHHC9, and CD274 expression, although the three-gene response score yielded an exact permutation P = 0.10. Independent assessments agreed on all 48 evidence ratings (quadratic-weighted κ = 1.000). Discussion Evidence gating distinguished reproducible molecular associations from complete cross-layer mechanistic support. The P. stomatis–glycocholate association, accompanied by complementary GCA-responsive host transcription, represents a candidate for further investigation rather than an established microbial–metabolite–host mechanism.