Generalizing Drug–Drug Interaction Risk Prediction to Chemically Novel Compounds with Bilinear Interaction Learning and Calibrated Hybrid Inference

Reliable drug–drug interaction (DDI) prediction is essential for polypharmacy safety and for prioritizing risky combinations during early-stage drug discovery. A persistent obstacle is inductive generalization: models that perform well under convenient i.i.d. splits often degrade when evaluated on chemically novel compounds. In addition, some feature sets used in prior work can inadvertently encode postmarket clinical proxies, complicating claims about preclinical utility. Here, we present BiG-Stack, a cold-start-oriented DDI prediction framework that combines (i) a gated multiview Siamese encoder over chemical fingerprints, protein targets, pathway-level biology, and 31 mechanistically motivated biological descriptors, (ii) an asymmetric bilinear interaction operator to model cross-feature dependencies between drugs, and (iii) a stacked hybrid inference layer that fuses deep interaction features with mechanistically motivated preclinical tabular descriptors to produce calibrated probabilities with interpretable attributions. We evaluate under a strict scaffold-disjoint single cold-start protocol in which every test pair contains at least one drug whose Bemis–Murcko scaffold is absent from training. On this inductive setting, BiG-Stack achieves AUROC = 0.957, AUPRC = 0.898, and F1 = 0.902. A preclinical robustness audit that removes postmarket clinical features from the tabular space preserves strong performance (AUROC = 0.946, AUPRC = 0.883, F1 = 0.882), supporting the model’s utility when only chemistry and preclinical biology are available. Finally, a time-split pharmacovigilance analysis links high-confidence predictions to elevated future FAERS disproportionality signals, providing complementary evidence of translational relevance. Together, BiG-Stack offers a leakage-resistant and interpretable approach for inductive DDI risk screening under chemical novelty.