The abstract is read at the publisher; the summary is JClub's.
BiomedCAT converts biomedical slide decks into RTX-KG2c subgraphs, achieving 0.71 accuracy on 517 audited mentions while recovering 81–94% of explicit English terms.
- Why it matters: Mechanistic ideas often start in figures and slides, but KG tools require graph vocabulary, creating a gap for traceable, user-shaped context extraction without external document sharing.
- What they did: The workflow uses a small vision-language model to describe slides, a 4-billion-parameter agent to map terms to Biolink classes with an existence gate, then retrieves identifiers via Translator or a local table and expands a profile-weighted two-hop context graph.
- The result: In expert-annotated decks, linking finds 65% of terms online and 26% fully locally, and in an FSHD case three DUX4-centered graphs overlap by 7–15% while missing 7 of 21 tested compounds from RTX-KG2c.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.