Chemical Information Modeling for Component-Level Review Prioritization of Adverse-Reaction Signals in...

Postmarketing safety signals for multicomponent medicines are usually assigned to a whole product, leaving unresolved which component chemistry should be reviewed first. We formulate this problem as chemical information modeling over a heterogeneous safety graph connecting Kampo formulas, crude-drug components, ingredients, protein targets, protein-protein interactions, and adverse-reaction terms. HerbPairIAM represents components, unordered component pairs, and adverse reactions as fixed target-centered graph profiles, then learns adverse-reaction-conditioned attention to score formula-reaction candidates and rank component-level review priorities. High-confidence formula-adverse reaction signals were curated from JADER and FAERS using a four-estimator disproportionality vote rule requiring at least three positive estimators and at least three coreports. In a leakage-controlled tested-universe benchmark containing 704 positives among 3965 candidate pairs, HerbPairIAM achieved an AUROC of 0.823 and an AUPRC of 0.517 across 30 matched folds, outperforming seven tabular and graph-based baselines. Ablations showed that the unordered pair branch and fixed graph-derived profiles drove the performance gain, and formula-held-out evaluation retained transfer to unseen formula compositions. Generalization was substantially weaker in the ADR-held-out setting, localizing the main transfer boundary to sparse adverse-reaction target annotation; the model is therefore more reliable for prioritizing within a represented adverse-reaction space than for cold-start adverse-reaction phenotypes. Calibration, review-threshold analysis, leave-one-component perturbation, and package-insert auditing linked high-scoring candidates to compound, target, and documentary evidence, with 192 of the top 500 tested nonsignal candidates supported by Japanese package inserts, an enrichment that attenuates after adverse-reaction-family matching and is best read as family level concentration of review candidates rather than within-family discrimination. Two compound-target structural follow-ups provided illustrative plausibility checks for selected compound-target axes rather than mechanistic or clinical-toxicity validation. The resulting scores are intended for safety-review triage and mechanism-focused follow-up, not for direct causal attribution of toxicity.