Pocket Definition Contract

A pocket definition contract states how the binding site is chosen before modeling or evaluation. Pocket definition is part of the task, not a harmless preprocessing detail, because it controls what information is available to the model.

$$B=\varphi (P,C)$$

where $P$ is the protein structure and $C$ is the allowed context for selecting the pocket.

Pocket Modes

Mode	Allowed context	Use for	Risk
known site	public annotation or prior biological knowledge	focused screening, known target workflow	annotation may be benchmark-specific
predicted site	pocket predictor output from receptor only	prospective or blind-ish workflow	predictor quality becomes part of method
ligand-defined	residues near a bound ligand	retrospective pose analysis	test ligand pose leakage
blind grid	whole protein or broad search region	blind docking	larger search space and lower baseline
template-derived	homolog or template complex	structure transfer	template and homolog leakage
user-defined box	manually specified center/radius	operational docking workflow	may hide unavailable expert knowledge

Required Fields

Field	Record
receptor source	experimental, predicted, template, ensemble, prepared structure
pocket source	annotation, predictor, ligand, template, manual box, blind grid
selection rule	residue cutoff, grid size, center, chain, cofactors, exclusions
inference availability	whether the same pocket rule is available without the test ligand
split interaction	whether pocket source leaks homolog, template, ligand, or assay information
evaluation impact	how failures in pocket definition are counted

Claim Boundary

Claim	Pocket requirement
pose recovery around known ligand	ligand-defined pocket can be acceptable if stated explicitly
prospective docking	pocket must be known, predicted, or defined without test pose
virtual screening	pocket rule must be fixed before candidate labels are used
structure-aware generation	conditioning pocket must match deployment information
new target generalization	pocket source must respect protein-family/template split

Checks

• Is the pocket chosen before seeing the test ligand pose?
• Would the same pocket be available for a novel ligand or target?
• Are failed pocket predictions counted or silently filtered?
• Does the pocket rule use benchmark-only annotations?
• Is the receptor structure source independent of the evaluated complex?

Related

• Pocket
• Structure-Based Modeling
• Docking workflow
• Protein-ligand split
• Template leakage