Pocket Representation

Pocket representation describes how a protein binding site is converted into model input. The pocket can be represented as residues, atoms, surfaces, grids, graphs, coordinates, or learned embeddings.

A distance-defined pocket around ligand coordinates $X_L$ can be:

$${\mathcal{P}}_r=\{i\in P:\underset{j\in L}{\min }\Vert x_i-x_j{\Vert }_2\le r\}$$

where $r$ is a cutoff radius.

The pocket representation is then:

$$z_{\mathcal{P}}=f_{\theta }({\mathcal{P}}_r,X_{\mathcal{P}},A_{\mathcal{P}})$$

where $X_{\mathcal{P}}$ are coordinates or features and $A_{\mathcal{P}}$ is optional adjacency or contact structure.

Common Representations

Representation	Captures	Main risk
residue sequence window	local sequence context	misses 3D proximity and long-range contacts
residue graph/contact map	residue-level geometry	cutoff and missing residues change graph
atom-level graph	local chemistry and coordinates	atom typing, protonation, and side-chain uncertainty
3D grid or voxel	spatial density around pocket	resolution, frame, and memory cost
surface patch	shape and chemical surface features	surface generation parameters affect features
learned embedding	compact representation for retrieval or prediction	hard to audit what information is preserved

Pocket Source Contract

Source	Available at deployment?	Claim boundary
known binding site	yes, if site is annotated	tests site-conditioned modeling, not blind site discovery
predicted binding site	yes, if predictor is part of workflow	includes predictor error in the pipeline
ligand-defined pocket	no for unknown binders	useful for retrospective analysis, risky for blind prediction
residue cutoff from reference complex	no for new pairs	can leak the answer if used in evaluation
whole-protein input	yes	shifts burden to model to find relevant region

If a pocket is defined from a bound test ligand, the task is no longer blind pocket discovery or deployment-like docking.

Key Risks

• Pocket extraction can leak ligand information if defined from the test ligand.
• Different cutoffs can change the apparent task.
• Missing side chains, cofactors, metals, and waters may alter interactions.
• Whole-protein context can matter even when the model only sees a local pocket.
• Pocket similarity across train and test can hide target-family leakage.
• Predicted pockets can fail silently if the benchmark only scores successful extractions.

Checks

• Is the pocket known, predicted, transferred, or ligand-defined?
• What atoms, residues, chains, waters, metals, and cofactors are included?
• Does the representation preserve rotation/translation invariance or equivariance?
• Is the pocket representation available at deployment time?
• Is pocket similarity controlled across train and test?
• Are failed pocket detections, empty pockets, and missing structures counted?
• Does the evaluation separate pocket detection quality from downstream scoring quality?

Related

• Pocket
• Binding site
• Protein-ligand interaction
• Equivariance
• Graph neural networks
• Protein family split