by Xiaoqing Ru, Yansu Wang, Quan Zou, Mengting NiuReliable generalization is essential for deploying drug–drug interaction (DDI) and drug–target interaction (DTI) models in drug discovery, pharmacovigilance, and clinical decision support. However, strong performance on random splits or weakly constrained evaluations may not reflect how well a model generalizes to new drugs, targets, or deployment conditions. In pairwise prediction, test instances may still resemble training examples in molecular structure, sequence, or network context. Such overlap can inflate performance estimates and make interpolation appear to be extrapolation. In this Perspective, we argue that cold-start evaluation should be treated as an auditable deployment claim. We organize common evaluation settings by their source and scope of novelty, ranging from random splits and entity-disjoint protocols to structure-level, temporal, domain-disjoint, and many-domain evaluations. The L0–L5 framework clarifies which deployment claim each setting is designed to test. Coverage, overlap, and distance quantify the resulting train–test discrepancy. We further provide a practical checklist for leakage control, validation alignment, candidate-pair construction, and domain-level risk reporting. This Perspective clarifies the boundaries and evidential scope of cold-start evaluation, providing a common basis for interpreting and comparing results across DDI and DTI studies.