Ligand-Free Is Not Closed: Why Occupancy Labels Rank a Degenerate Predictor First
Abstract
Cryptic pockets are the main route to targets judged undruggable, and the conformation a predictor emits without a ligand decides which pockets a pipeline can find at all. Auditing that means reading a ligand-free structure’s open/closed state. We show this evaluation cannot be validated, because the label it rests on does not encode the state. Taking steric admissibility of the cognate ligand as ground truth, 52% of ligand-free holdouts exclude it, 17% are intermediate and 31% admit it outright, so crypticity is a property of a structure, not of a protein. Three geometrically independent readouts consequently sit at 54–64% on ligand-free holdouts against 85–92% on ligand-bound ones, each at its label’s 69% ceiling, not at its own. Worse, the label inverts the ranking it exists to produce: a degenerate always-closed predictor outscores a per-item perfect one by 48 points. We give a continuous label-free steric readout, robust to its one free parameter, and show on a synthetic predictor ladder that it removes the inversion. Evaluation arms must also be scoped to the training cutoff (18% mis-assignment, zero direction reversals).