Project 4: Drug-label embedding clusters¶
Mentors: To be announced
Problem: Drugs are officially grouped by pharmacological class, but it's unclear whether their warning-label language actually clusters that way, or clusters instead around shared risk profile.
Context: Built on DailyMed FDA drug label text (warnings and indications sections), anchored to ML4LLM Ch.3 · proj8: All to all cosine similarity (helper).
Goals: Do drugs with black-box warnings cluster together in embedding space, independent of their pharmacological class?
Deliverables: A notebook that embeds each label's warnings/indications section, builds the all-to-all cosine similarity matrix, visualizes it as a heatmap/dendrogram, and overlays drug class and black-box-warning status to see which grouping the embedding space actually reflects.
Showcase: TBD
References:
- Multimodal Model with Text and Drug Embeddings for Adverse Drug Reaction Classification: combines drug-label text embeddings with structured drug representations to study shared risk patterns, the same embedding-space grouping question this project asks of DailyMed warning text.