Chemical Motifs Associated with FAERS-Derived Severe Cutaneous Adverse Reaction Disproportionality Signals: An Interpretable Pharmacovigilance-Driven Cheminformatics Study

Severe cutaneous adverse reactions (SCARs) are rare, life-threatening drug hypersensitivity syndromes. Although pharmacovigilance can identify drugs disproportionately reported with SCARs, it does not reveal which local chemistries recur among them. To address this, we assessed whether drugs with FAERS-derived SCAR disproportionality signals share interpretable chemical motifs. We screened FAERS data from 2004Q1 to 2024Q3, identified 5523 drugs with available Simplified Molecular-Input Line-Entry System (SMILES) representations, and constructed a signal-enriched dataset of 1676 compounds with nominally significant broad-SCAR associations after excluding predefined therapeutic/supportive confounders. Compounds were assigned to positive-signal [natural logarithm of reporting odds ratio (lnROR) > 0, n = 1219] or non-positive-signal (lnROR ≤ 0, n = 457) classes and encoded with 9753 explicitly mappable atom-centered local substructure descriptors. A LightGBM signal-classification model evaluated using random repeated nested cross-validation (six-fold outer × 50 repeats) achieved moderate internal discrimination (mean area under the receiver operating characteristic curve = 0.7041 ± 0.0337). Descriptor-space cluster-based repeated nested cross-validation, designed to reduce train–test structural leakage, yielded lower but still above-chance performance (mean ROC AUC = 0.6409; permutation p = 0.001), indicating that random-split estimates should be interpreted as optimistic for structurally novel compounds. Sensitivity analyses using minimum SCAR case-count thresholds and retention of predefined therapeutic/supportive drugs showed broadly similar performance and motif rankings. SHapley Additive exPlanations (SHAP) analysis revealed a fragment-level contrast: allylamine-like, ethanolamine-related, and diaminopropane-related motifs were associated with higher positive-signal class probability, whereas phenol and pyrimidine motifs were associated with lower positive-signal class probability. These findings suggest that FAERS-derived broad-SCAR signal direction is not chemically random within the selected dataset. Overall, the proposed framework should be viewed not as a direct predictor of absolute clinical SCAR risk but as an exploratory, pharmacovigilance-driven cheminformatics approach for prioritizing compounds and motif families for further SCAR-focused evaluation.