PharmaLink-DB: A Regulatory-Derived Knowledge Graph for Safety-Aware Biomedical Recommendation

PharmaLinkDB is a multi-relational knowledge graph linking drugs and diseases, built automatically from FDA regulatory sources. It jointly encodes positive (therapeutic indication) and negative (contraindication) clinical evidence in a single graph, along with two auxiliary similarity relations used to densify the entity neighbourhoods. Reliable identification of therapeutic indications and contraindications for approved drugs is central to clinical pharmacology, yet the two evidence types are almost always modeled separately: a drug's therapeutic uses and its documented harms rarely sit in the same embedding space, and existing resources are typically static, versioned snapshots rather than sources that stay current with regulatory updates. Existing biomedical knowledge graphs also tend to face familiar problems sparse and noisy coverage of any one relation type, entity vocabularies that don't normalize cleanly to canonical ontologies, and no built-in mechanism for the positive and negative evidence to be used differently at inference time despite carrying very different clinical stakes. PharmaLink-DB addresses this gap. It is a multi-relational knowledge graph linking 755 drugs and 697 diseases across 13,489 triples and four relation types, built automatically from FDA DailyMed Structured Product Labels and FAERS adverse-event cross-references over a rolling three-month ingestion window. The graph jointly encodes 3,601 is_indicated and 1,367 is_contraindicated triples: positive and negative clinical evidence in one shared entity space; densified by two auxiliary similarity relations: 1,788 drug-drug structural-similarity edges (Tanimoto coefficient over Morgan/RDKit fingerprints) and 6,733 disease-disease ontological-similarity edges (Jaccard coefficient over shared Disease Ontology ancestor terms). All entities are normalized to canonical identifiers ChEBI or DrugBank for drugs, Disease Ontology or Orphanet for diseases: with both the ID and the human-readable label carried in every row, so the data loads directly as labeled triples or joins against external ontologies without a separate lookup table. Unlike prior work that collapses indication and contraindication evidence into one relation or trains them in disconnected embedding spaces, PharmaLink-DB's schema is built to support a rank-then-veto inference pattern, where candidate diseases are ranked by indication score and any candidate meeting a data-calibrated contraindication threshold is excluded regardless of rank; evaluated in the companion paper against four representative KGE architectures (TransE, DistMult, ComplEx, RotatE) and validated against the Hetionet benchmark under an identical protocol. We hope PharmaLink-DB's regulatory grounding, joint positive/negative evidence structure, and update-friendly construction pipeline support further work in KG-based drug repurposing, pharmacovigilance, and safety-aware biomedical recommendation. Files File Description unified_records.json.zip MedJsonify-normalized intermediate records, one per DailyMed SPL / FAERS cross-reference, prior to triple extraction. triples_indicated.csv indicated edges extracted from SPL Indications and Usage sections. triples_contraindicated.csv contraindicated edges extracted from SPL Contraindications sections, cross-referenced against FAERS. triples_drug_similarity.csv structurally_similar edges from Tanimoto similarity over Morgan/RDKit fingerprints (τ ≥ 0.7). Symmetric — both directions included. triples_disease_similarity.csv ontologically_similar edges from Jaccard similarity over shared Disease Ontology ancestor terms. Symmetric — both directions included. kg_full.csv All four relations concatenated into a single file, ready to load. CHANGELOG.txt Version history across rolling three-month extraction windows. Relations Relation Abbreviation Semantics is_indicated CtD Compound–treats–Disease is_contraindicated CcD Compound–causes–Disease is_drug_similar CrC Compound–resembles–Compound is_disease_similar DrD Disease–resembles–Disease