RxPairEvid
Description
RxPairEvid is a license-clean, machine learning ready dataset of 50,000 drug--drug pairs with pair-level pharmacovigilance context from the FDA Adverse Event Reporting System (FAERS).
Files
Steps to reproduce
Download the files from this record (ddi_pairs_50k.csv, schema.sql, codebook.md, audit_subset_signal_quantiles.csv, audit_subset_strata_counts.csv, provenance.md, checksums.txt) and verify integrity with the SHA-256 checksums. Load the CSV with your tool of choice (e.g., pandas) or create tables in PostgreSQL using schema.sql and then import the CSV. To rebuild from sources: ingest FAERS DRUG/REAC, normalize names and map drugs to InChIKey14 (IK14); enumerate co-medication pairs and join MedDRA Preferred Terms by PT code; for each pair–PT compute 2x2 counts with +0.5 continuity correction; compute PRR, ROR, and the log-scale 95% lower CI; roll up signals under strict floors (a_raw >= 3, pair >= 10, pt >= 10) to get per-pair maxima, coverage, and the best PT code; stratify-sample 50k rows by ROR95_LCL, PT coverage, and ATC level, deduplicate by sorted IK14 pair key, and export the CSV plus audits. Third-party payloads (DrugBank, KEGG, PDBbind, MedDRA text) are not redistributed; schema.sql shows where they attach if a user has licensed access.