Warm-start evaluation tests combinations whose compounds and proteins may have appeared during training, while withholding the specific compound–protein pairs. It measures whether a model can predict new relationships among known entities. section 11.5.5
This setting reflects practical drug-discovery tasks such as predicting known drugs against new targets or new drugs against known targets. Because repeated entity exposure can support memorization, warm-start scores may overstate performance on genuinely new chemistry or targets. section 11.5.5
Compare warm-start evaluation with stricter alternatives:
- Cold-drug splits hold out entire drugs; cold-target splits hold out proteins; cold-drug/cold-target splits hold out both partners. section 11.5.5
