MicrobExpert: A mixture-of-experts framework for cross-domain knowledge mining of food–microbe–disease associations from biomedical texts

The gut microbiome is vital to human health, with diet as a primary modulator. Understanding food–microbe–disease interactions is essential for advancing personalized nutrition and disease prevention, yet extracting structured knowledge from biomedical and food science literature remains challenging due to semantic, cross-domain, and resource limitations. We propose MicrobExpert, a Mixture-of-Experts (MoE) framework integrating multiple BERT-based models with a gated routing mechanism to enhance RE accuracy, domain adaptability, and scalable nutrition-related knowledge base expansion. To effectively identify and locate candidate entity pairs and context segments, we introduce a named entity recognition (NER) module based on dictionary matching and integrate it with MircobExpert, forming a unified text mining pipeline for food–microbe–disease associations.Experimental results on a manually curated gold standard corpus (GSC) demonstrate that MicrobExpert achieves superior extraction accuracy and robust domain adaptability compared to strong baselines. Using this pipeline, we constructed two silver standard corpora (SSC), including 9577 disease–microbe and 3808 food–microbe associations, significantly enriching the knowledge base. We also developed a microbiome-mediated knowledge graph to facilitate downstream applications such as personalized nutrition, microbiome-guided disease management, and automated knowledge graph construction. The constructed knowledge graph, including complete datasets and reproducible visualization configurations, is publicly available at: https://github.com/Gsx2025/food-microbe-disease .