Published papers hold more assay data than the databases
An AI workflow lifts results out of papers and builds a larger, versioned database for protein degradation drugs.
Picture made for ABUREAU.
Much lab data is trapped in the text of papers. A new preprint shows a workflow that extracts it into a structured database for targeted protein degradation, a way of designing drugs.
The method uses a language model, with a small set of expert annotations to refine its prompts. On molecular glue papers it reached a record-level F1 score of 0.98 using seven annotated publications. Applied to open-access literature, it added 4,483 records from 598 publications.
The authors say their database holds 81% and 92% more molecular glue and PROTAC records than the standard reference sets. Expert audits found 92 of 100 and 33 of 40 new records correct. The public version, AutoTPD+, is versioned for monthly updates, and models trained on it showed 8.9 to 18.2% lower error than models trained on the older PROTAC-DB data.
For people who own or use data, the lesson is that a curated, current, structured set can beat the established one. Data sitting in documents is a supply waiting to be organised, and the audit rate shows what a buyer should ask for.
