Cold-start evaluation tests whether a drug–target affinity model generalizes to entities absent from training data. Unlike warm-start evaluation, it prevents the model from seeing the same compound or protein in training and testing, making genuinely new chemistry or targets harder to predict. section 11.1.1 section 11.5.5
Use distinct holdout designs to control which entities remain unseen:
- Cold-drug splits hold out entire drugs, so every test drug is absent from training. section 11.5.5
- Cold-target splits hold out entire proteins, so every test protein is absent from training. section 11.5.5
- Cold-drug/cold-target splits hold out both partners in every test pair. section 11.5.5
- Compare these results with warm-start scores cautiously, because repeated exposure to compounds and proteins can support memorization and overstate performance. section 11.1.1 section 11.5.5
KIBA provides a realistic motivation: its sparse interaction matrix resembles incomplete experimental data, while its diverse compounds make prediction for genuinely new chemical space difficult. Cold-start evaluation therefore tests a stricter form of generalization than predicting unseen combinations of familiar drugs and proteins. section 11.1.1 section 11.5.5
