Drug-centric knowledge graph: representing therapeutic modality, pharmacology and multiscale biology

Abstract

Biomedical knowledge graphs link targets to mechanisms, diseases and phenotypes, but target-centered representations omit much of what distinguishes therapeutic interventions. We introduce a drug-centric knowledge graph, whose identity model, schema and information-selection priorities are organized around the therapeutic intervention, and a graph built on this design. Each intervention has one identity, linked through relation types shared by all compounds, including target actions, indications, adverse effects, classifications, half-life and direct links to pathways, Gene Ontology concepts, cells and anatomy, and through modality-specific relation types describing small-molecule chemistry, antibody sequences and formats, protein-therapy sequences and antibody-drug conjugate components. Integrating 38 source resources, the graph contains 166,261 nodes of 27 types, 8,106,906 directed edges of 129 relation types and 18,429 compounds in eight therapeutic categories. Construction combines source-specific extraction, compound-identity resolution, language-model-assisted vocabulary and predicate mapping, confidence scoring and connectivity filtering; an expert-reviewed workflow adds investigational therapeutics without altering the base graph. The design offers a reusable template for representing interventions as structured, multimodal objects.