How do we extract and link relationships from scientific research?

We extract relationships from scientific research by using large language models (LLMs) to scan studies and pull out each of the components reported (i.e., variables, and the following as available: mechanism type, statistic type, statistic value, confidence interval, and p-value). For every extraction, traceability to the original source is preserved by storing all data at the highest resolution, true-to-source, along with links to the underlying sources, which are provided to users.

Once extracted, we link findings together by mapping each of the variables in a relationship to a higher-level concept or topic grounded in external ontologies such as MESH, ICD-10, SNOMED CT, and DSM-5, LNC, and Wikidata. For example, a variable named “daily aspirin usage” would be assigned to the topic of “Aspirin” [MESH ID: D001241, Wikidata ID: Q18216], and thus discoverable by searching for that associated topic. The result is an up-to-date and interoperable graph of evidence-based findings.