Consistency-regularized graph-tabular evidence fusion for cold-start drug-drug interaction prioritization

Background Drug-drug interactions (DDIs) remain difficult to prioritize for new, rarely prescribed, or poorly characterized drugs because interaction evidence is sparse and conventional random pair splits can overestimate generalization. Objective This study evaluated whether a consistency-regularized heterogeneous graph-tabular model can improve cold-start DDI prioritiza-tion under a strict drug-disjoint protocol. Methods We constructed RxPairEvid, a multi-source evidence resource that links curated drug knowledge, drug-protein-pathway relations, protein interaction context, side-effect resources, transcriptomic perturbation evidence, and spontaneous-report safety signals. RxDDI-ConsisGAT encodes mechanistic context with relation-aware GATv2 layers, encodes pairwise evidence through a tabular pathway, and fuses both views for binary interaction scoring. Training uses two stochastic forward passes with independent dropout masks and penalizes probability disagreement alongside label-smoothed binary classification loss. Results On held-out cold-start testing, RxDDI-ConsisGAT achieved AUROC = 0.8352 and AUPRC = 0.6910. Same-protocol baselines and ablations showed lower test AUROC for the no-consistency model (0.8255), tabular-only model (0.7734), graph-only model (0.7163), HeteroSAGE-fusion model (0.7318), logistic regression (0.7597), and HistGradientBoosting (0.7803). Diagnostic analyses showed useful score separation, broadly reasonable calibration, and interpretable evidence patterns. Conclusion The findings support graph-tabular evidence fusion as a practical strategy for expert-facing cold-start DDI triage, while prospective validation and threshold calibration remain necessary before clinical deployment.