Knowledge Graphs for Drug Discovery Evidence Management
Holding
A drug discovery knowledge graph should not be a decorative map of entities. It should be an evidence management system that records claims, provenance, confidence, contradiction, and experimental context.
Authority
Scientific software quality, research integrity, and AI risk management all require that conclusions be traceable to their evidence. When a graph feeds candidate selection or model training, missing provenance becomes a material engineering risk.
Issue
The failure mode is graph laundering. A weak association enters the graph from a paper, database, or model output, then appears later as an established biological relation. The graph has increased confidence by repetition rather than evidence.
Resolution
The graph should store source, method, assay, organism, tissue, disease context, date, confidence, and negative evidence. Edges should distinguish observed fact, inferred relation, curated assertion, and model-generated hypothesis.
Evidence Package
The record should include ingestion provenance, curation rules, ontology versions, conflict resolution, evidence grades, model-use logs, and decision exports. A discovery graph is mature when it can explain why a target or compound became plausible.