Outside the Training Map

A model's competence has a boundary set by the examples it was trained on. Inside that boundary, its predictions are reliable. Outside it, the model still produces a confident number, but that number is an extrapolation with no supporting evidence. Drug discovery deliberately searches for compounds and targets that are unlike anything already known, so the most valuable cases are often the ones furthest outside the training distribution. This is why a model can look excellent on held-out test data and still fail on the first genuinely novel candidate.