A pocket is a local region of a protein that can host a ligand, cofactor, substrate, or designed molecule. In structure-based modeling, the pocket is often the practical input object rather than the full protein.
Modeling Views
- Residue set around a known ligand or binding site.
- 3D grid, surface patch, atom graph, residue graph, or point cloud.
- Conditional context for docking, pose generation, and scoring.
Given protein coordinates and a ligand reference position , a simple distance-based pocket can be defined as:
where is a cutoff radius.
Pocket Source
Pocket definitions differ by available evidence:
A ligand-defined pocket is valid for retrospective pose analysis, but can leak answer information if the intended deployment setting must find the binding site without a bound ligand.
Representation Boundary
The pocket can be represented as:
- residues within a cutoff;
- atoms within a cutoff;
- surface points and normals;
- voxel grid;
- residue or atom graph;
- local coordinate frame;
- sequence window around annotated residues.
Each representation makes different assumptions about rigidity, solvent, protonation, side-chain flexibility, and coordinate-frame availability.
Leakage Risk
Pocket extraction should be part of the protocol:
If the extraction function uses the test ligand pose, bound ligand center, or benchmark-only annotation that would not be available in screening, the task becomes easier than the claimed deployment setting.
Checks
- Is the pocket defined from a known ligand, predicted binding site, or whole-protein scan?
- Does the definition include backbone atoms, side chains, waters, metals, or cofactors?
- Does pocket extraction leak test-set ligand information?
- Is the pocket rigid, flexible, or updated during modeling?
- Is the same pocket extraction protocol used for train, validation, test, and inference?
- Are apo and holo structures handled consistently?
- Are local coordinate frames defined without using unavailable ligand information?