Named entity recognition identifies the players in biomedical text. Relation extraction identifies the relationships between them — and those relationships are where the scientific value resides. A knowledge graph that contains metformin and type 2 diabetes mellitus as isolated nodes is less useful than a spreadsheet. A knowledge graph that asserts metformin has-indicated-use type-2-diabetes-mellitus, metformin inhibits hepatic gluconeogenesis, and metformin is-contraindicated-in renal-failure is a navigable model of pharmaceutical knowledge that supports clinical decision support, drug discovery, and safety monitoring.
The Relation Types That Matter Most
For pharmaceutical and clinical knowledge graphs, the most valuable relation types to extract are: has-indicated-use (drug treats condition), has-adverse-effect (drug causes adverse event), has-contraindication (drug should not be used in condition), has-mechanism (drug acts via pathway or target), interacts-with (drug modifies the effect of another drug), and is-biomarker-for (molecular feature predicts condition or response). Each of these relation types has a different linguistic signature in text and requires different extraction strategies.
Extraction Approaches
Rule-based extraction using syntactic patterns is still competitive for well-defined relation types in structured text such as drug labels and summary of product characteristics documents, where the language is highly constrained. For clinical notes and research publications, machine learning approaches — particularly sequence-to-sequence models that jointly perform entity recognition and relation extraction — outperform rule-based systems on recall while maintaining acceptable precision. A hybrid approach that applies rules for high-confidence cases and machine learning for ambiguous ones achieves the best precision-recall trade-off for most production applications.
Populating and Validating the Graph
Relation extraction at scale produces large volumes of candidate assertions that must be validated before entering a curated knowledge graph. A tiered validation approach — high-confidence automatic assertions, medium-confidence assertions requiring light expert review, low-confidence assertions queued for manual curation — allows the knowledge graph to grow continuously while maintaining the quality standards required for regulated downstream applications. Provenance tracking, which records the source document and extraction model version for each assertion, is essential for defensibility and for updating the graph when source documents are revised or retracted.
