Abstract
Diet profoundly influences gut microbial composition and metabolism, yet mechanistic pathways linking dietary exposures to human health remain fragmented across the literature. To systematically organize and integrate this evidence, we constructed an evidence-weighted Diet-Microbiome-Health Knowledge Graph (DMH-KG) from 1,309 curated PubMed abstracts (2023-2025), using a standardized schema spanning 11 entity categories and 12 relationship types with explicit polarity labels. Entities and relationships were identified through manual annotation, yielding 10,270 entity mentions and 4,866 relationships. Expert-guided consolidation of synonymous and lexically variant terms, along with pruning of disconnected concepts, resulted in 4,766 unique entities connected by 4,772 polarity-weighted edges. To prioritize robust biological signals, a composite edge-weighting function was applied, integrating both relationship frequency and polarity. Network analysis revealed a modular small-world structure centered on microbial and inflammatory mediators. High-confidence pathways emerged, including the probiotic-SCFA-immunity axis (34 supporting documents) and the high-fat diet-LPS-endotoxemia cascade (8 documents), both of which are central to age-related immune modulation and metabolic health. Quantitative validation against five KEGG and Reactome pathways demonstrated high biological fidelity: the DMH-KG recovered all reference diet-microbiome-outcome edges for Butanoate metabolism and Secondary Bile Acid biosynthesis (100% coverage) and achieved a mean pathway-level entity coverage of 92.0%, measured as the proportion of predefined pathway components represented in the graph. A comparative pilot study further demonstrated that DMH-KG augmentation improves mechanistic specificity inference across Diet-Microbiome-Health interactions. These features position the DMH-KG as a scalable platform for mechanistic inference across Diet-Microbiome-Health interactions, with direct relevance to immune regulation, muscle health, metabolic aging, and chronic disease prevention. The framework preserves evidence provenance, relationship polarity, and biological direction, supporting both discovery science and AI-driven nutritional reasoning.