Wiki · Bio × AI Methods · concept
Knowledge graph
A network of biomedical facts — genes, drugs, diseases, proteins linked by typed relationships you can query and reason over.
A knowledge graph stores facts as a network, not as prose. Each entity — a gene, a drug, a disease, a protein — is a node, and each fact connecting them is a labeled edge: treats, targets, causes, interacts with. “Aspirin inhibits COX-1” isn’t buried in a sentence; it’s a typed link between two nodes that a machine can follow. The result is biology written as a structure you can query, not a document you have to read.
How it works
Nodes and their labeled edges are assembled from two sources: curated databases (drug–target tables, disease ontologies, pathway maps) and relationships mined from the literature at scale. Once the graph exists, you use it in two ways. You can reason by traversal — walk a path from drug to target to pathway to disease to explain why a compound might work — or you embed the graph, turning nodes into vectors (see embeddings) so a model can predict edges that aren’t yet drawn. That link-prediction step is how a KG proposes new connections. Increasingly the graph is paired with a language model that retrieves the relevant subgraph and puts the answer into words.
Why it matters (for bio × AI)
A knowledge graph is the structured, auditable memory behind agentic drug discovery. When an agent proposes a target or a repurposing candidate, the KG lets it surface a readable evidence chain — this drug, this protein, this pathway, this disease — instead of a black-box score you have to trust on faith. It also grounds a large language model: rather than letting the model free-associate and hallucinate a citation, you anchor its answers to facts that actually exist in the graph. This is the backbone of the agentic-drug-discovery chapter — the difference between a confident guess and a claim you can trace.
Related
See large language model for the reasoning engine often bolted onto a graph, and embeddings for how a graph gets turned into vectors for link prediction.