Knowledge Graph
A structured representation of entities and the relationships between them — used in biomedical databases, literature mining tools, and AI systems that need to reason over facts rather than text.
What it means
A knowledge graph is a network of entities (people, concepts, genes, drugs, papers) connected by typed relationships (treats, interacts with, co-authored, is-a). Each connection is a triple: subject → predicate → object. For example: metformin → treats → type 2 diabetes, or TP53 → regulates → CDKN1A.
Knowledge graphs differ from plain text in a crucial way: the relationships are explicit and queryable. You can ask “which drugs target proteins in the p53 pathway?” and get a precise answer by traversing the graph — something that requires imprecise natural language search over plain text.
Major scientific knowledge graphs:
- UniProt (protein functions and interactions)
- STRING (protein-protein interaction network)
- ChEMBL (drug-target interactions and bioactivity data)
- PubChem (chemical structures and biological activities)
- Hetionet (integrates multiple biomedical databases into a heterogeneous network)
- SPOKE (the Scalable Precision Medicine Oriented Knowledge Engine)
Knowledge graphs and AI
Retrieval-augmented generation (RAG) with knowledge graphs: Rather than retrieving text chunks, some RAG systems retrieve subgraphs — structured facts that are then fed to an LLM as context. This gives the LLM more precise, structured information and reduces hallucination on factual claims.
Graph neural networks (GNNs) on knowledge graphs can predict missing links — for example, predicting which drug might interact with a newly characterized target by learning from the pattern of known interactions in the graph.
Literature-derived knowledge graphs: Tools like PubTator Central and INDRA automatically extract biomedical relationships from the literature and add them to knowledge graphs. This semi-automated extraction has errors, but at a scale (millions of papers) no manual curation can match.
Relevance for researchers
Literature tools like Semantic Scholar and Connected Papers are effectively specialized knowledge graphs — entities are papers, relationships are citations. When you explore a citation network, you are traversing a knowledge graph.
Drug-target interaction databases (ChEMBL, DrugBank) are knowledge graphs that AI drug discovery tools query to identify known associations, filter out already-known interactions, and validate novel predictions.