RxPairEvid: an auditable, machine-learning-ready dataset of drug–drug pairs with pharmacovigilance signal features and MedDRA PT-code rationales

Data sources

Redistributed source and derived outputs:

Referenced as optional attachment points (not redistributed in RxPairEvid-50 K):

Processing environment and reproducible schema

The verified internal build ran on PostgreSQL 14.19 (Ubuntu) with a psql 18.0 client. The exported schema report records the database layout (schemas stg, core, features, labels, ml, ref ) and provides row estimates and column types for reproducibility. This build has 11,521 canonical drugs as core.drug, 1,073,256 canonical drug pairs as features.pair_features_all and 50,340 pairs satisfying strict floors (a_raw≥3, pair≥10, PT ≥ 10) as seen in strict FAERS rollup table.

Canonical identifiers

Every drug has a standardized identifier (IK14 identifier, the first 14 characters of the InChIKey) which can be used to stabilize joins across sources, and to enable decent leakage-aware evaluation splits. Pairs are represented in unordered keys (A < B) to eliminate duplication and ambiguity in ranking.

FAERS ingestion, normalization, and mapping quality control

FAERS DRUG quarterly and REAC quarterly records are broken out into normalized staging tables, with powerful management of typical encoding problems. Mentions of drugs are normalised (case folding, removal of punctuations, removal of dosage/form/route tokens) and compared to IK14 with token-based dictionaries and database indexes. In the verified build, there are 322,635 distinct normalized names in the normalized FAERS name table, and the final FAERS name-to-drug mapping table has 59,507 mapped names, which gives it an auditable interface to quality checks of the mapping (e.g., frequency-ranking inspection of unmapped names).

Disproportionality statistics and strict vs loose regimes

We construct a 2 x 2 contingency table and calculate measures of disproportionality of each (drug A, drug B, PT). PRR is widely employed in signal generation of spontaneous-report data.² ROR is as well a popular tool and application of the lower confidence bound will assist in minimizing the false positives based on the limited number of counts.³ We apply a Haldane–Anscombe (+0.5) continuity correction before computing log-scale standard errors:

It has two regimes of evidence, coverage-oriented loose regime, and a conservative strict regime. Floors on minimum counts (a raw>3, pair>10, PT > 10) are imposed by the strict regime to eliminate instability in small counts and enhance interpretability of the retained signals.

Per-pair rollups and rationale pointer

Pair-level features are produced by rolling up eligible PT rows and retaining maxima and coverage counts:

The PT-code pointer enables audit-friendly interpretation without redistributing MedDRA PT text.

Public subset construction (50,000 rows)

RxPairEvid-50 K is deterministic stratified sample of strict FAERS rollup matrix. When there is a coarse ATC grouping; (i) ROR95 LCL bins, (ii) PT coverage bins, and (iii) a coarse ATC grouping are used to define the strata. The pairs of the stratum are ordered deterministically on the basis of md5(pair_id) and a predefined quota is sampled to obtain 50,000 rows. This enables the selection to be reproduced in rebuilds with the same strict matrix and strata definition.

Data records

The information is stored on Mendeley Data¹⁸ and is available in flat files. Table 1 contains the major fields in the primary CSV (ddi_pairs_50k.csv), and Table 2 contains the files that were added in the Mendeley record. Figure 1 gives an overview in high level of the RxPairEvid-50 K generation and release bundle, Figure 2 gives details of the steps of the FAERS processing and roll-up on a strict regime prior to sampling, and Figure 3 summarizes the internal database integration and optional attachment points of licensed resources.

Data validation

Validation is more about auditability and integrity, and not about predictive performance. They have been released with the sampling strata counts, signal quantile summaries, and SHA-256 checksums so they can be checked with the benefit of both integrity verification and rapid sanity checks. Moreover, high floors and confidence-bound screening minimize the small denominator instability, and rationale pointers are used to justify manual inspection of high-signaling pairs.

Reuse notes and limitations

Signals derived by FAERS are associative and prone to reporting bias, confounding, and stimulated reporting and should not be used as causal data of a DDI. We suggest RxPairEvid-50 K since (i) it is a benchmarking dataset to signal-based modeling (ii) it is an evidence layer used to construct labels in multi-evidence pipelines (iii) it is a transparent baseline used in ablation analysis. To evaluate leakage-aware, split on the drug level (IK14-disjoint) and then form pair splits, and use PR-AUC as a chief measurement in class imbalance.