Publication:

ADR•X: an interpretable, leakage-aware machine learning framework for sertraline adverse drug reaction signal detection using FAERS pharmacovigilance data

Adverse drug reactions (ADRs) are among the leading causes of preventable patient harm globally, and while the FDA Adverse Event Reporting System (FAERS) offers the most comprehensive post-marketing safety repository available, most published machine learning (ML) studies that work with this database introduce information leakage by incorporating outcome-derived disproportionality metrics—proportional reporting ratios (PRR) and reporting odds ratios (ROR)—directly as model features, thereby inflating performance estimates and undermining real-world generalisability. This study presents ADR•X, a LightGBM-based, leakage-aware framework designed to detect sertraline ADR signals from FAERS data using an approximately 208-variable feature space spanning patient demographics, physicochemical molecular descriptors, pharmacogenomic indicators, biology-guided multi-omics proxy variables, and mechanistic interaction terms, with all PRR-, ROR-, and frequency-derived variables explicitly excluded. Two model configurations were evaluated: an unweighted baseline and an inverse class-frequency-weighted variant. The baseline achieved an AUC-ROC of 0.53–0.54 and the imbalance-adjusted model reached 0.55–0.56. Global SHAP analysis identified dose mg, metabolic overload score, and polypharmacy flag as the three most influential predictors, while all remaining features clustered near zero, confirming the absence of leakage-driven dominance. The framework was deployed as a reproducible Streamlit research portal and is intended exclusively for population-level hypothesis generation, not individual clinical risk prediction. Modest AUC values reflect the bounded information content of voluntary reporting systems and represent honest signal estimation rather than model inadequacy. ADR•X demonstrates that biologically plausible and interpretable ADR signal detection is achievable from FAERS data without sacrificing methodological integrity.