Beyond unseen entities: Auditable cold-start evaluation for drug interaction prediction.

The L0–L5 framework makes cold-start evaluation an auditable deployment claim by quantifying train–test novelty and discrepancy using coverage, overlap, and distance.

• Why it matters: Random or weakly constrained splits can overstate drug–drug and drug–target interaction performance when test pairs resemble training data in structure, sequence, or network context.
• What they did: The Perspective organizes evaluation protocols by source and scope of novelty (e.g., entity-disjoint, structure-level, temporal, domain-disjoint, many-domain) and links each to a specific deployment claim.
• The result: It provides metrics and a leakage-control checklist to align validation and candidate-pair construction, enabling clearer interpretation and comparison of DDI/DTI cold-start evidence.

AI summary of the abstract · may contain errors · not medical advice · read the paper.