Metabolite-centric Knowledge Graph Unveils Predictive and Mechanistic Insights through Self-supervised Multimodal Learning

Abstract

Metabolomics research is advancing toward a comprehensive view of metabolites and their biological functions, yet conventional, dictionary-like databases fall short of presenting these complexities. To overcome this limitation, we developed MetaKG, a metabolite-centric knowledge graph that integrates diverse chemical, biological, and semantic data, coupled with a global self-supervised multimodal learning framework that generates advanced embeddings (MetaKGE) for both predictive and mechanistic analyses. MetaKG outperforms existing models in standard representation learning metrics and across 12 external datasets. Furthermore, MetaKG refines existing database information by detecting misannotated labels and uncovering previously overlooked biochemical relationships. Its predictive capability naturally yields testable biochemical hypotheses, as demonstrated in studies of G6PD deficiency and colorectal cancer. By transforming passive, database-scale information into extractable knowledge and actionable insights, MetaKG paves the way for hypothesis-driven discovery in metabolomics and holds promise for informing future research in clinical biomarker identification and therapeutic development.