GRAIN: Molecules Are Not the Right Granularity -- Active-Ingredient Modeling for Safe Medication Recommendation

GRAIN argues that active ingredients are a more appropriate representation for safe medication recommendation than either whole drug codes or molecular substructures. It normalizes medication codes through RxNorm and jointly models a drug-level DDI graph, an ingredient-level DDI graph, and an EHR-derived co-prescription graph. A selective state-space backbone handles long, irregular patient trajectories, while a proportional controller adjusts the accuracy-safety trade-off using the observed validation DDI rate. On MIMIC-IV under matched preprocessing, vocabulary, splits, and evaluation code, GRAIN improves Jaccard from 0.4488 to 0.4983, PRAUC from 0.6911 to 0.7485, and F1 from 0.5989 to 0.6453 versus a re-implemented MambaHealth baseline, while reducing drug-level DDI rate from 0.1875 to 0.0948.