Protein-Ligand Complex

A protein-ligand complex combines a protein binding site, ligand conformation, and interaction geometry. It is a central object in structure-based modeling.

A complex can be viewed as:

$$C=(P,L,X_P,X_L,I)$$

where $P$ is the protein or pocket, $L$ is the ligand identity, $X_P$ and $X_L$ are coordinates, and $I$ is the interaction context used by a model or evaluator.

Questions

• Is the pose geometrically plausible?
• Does the scoring function measure pose quality, binding affinity, or both?
• Which representation best preserves geometry and chemistry?
• Is the task pose generation, pose ranking, affinity prediction, or virtual screening?

Example Unit

A complex example is usually a pair plus geometry:

$$e=(P,L,X_P,X_L,y,s)$$

where $y$ is an optional label and $s$ is the source or protocol. The label may describe pose quality, assay activity, binding affinity, or benchmark membership. These should not be mixed without naming the target.

Geometry and Label Boundary

For docking and SBDD, geometry and assay evidence are different:

$$\text{pose quality}\ne \text{binding affinity}\ne \text{virtual-screening rank}$$

A pose can be geometrically plausible but biologically weak. A ligand can be active in an assay while a generated pose is wrong. Notes should state which claim is being evaluated.

Split Policy

Complex-level splits need both sides:

$$split\_key(C)=(protein\_family(P),ligand\_scaffold(L),\mathrm{source}(C))$$

For stronger claims, hold out both protein families and ligand scaffolds. Holding out only row IDs is usually too weak for structure-based generalization.

Checks

• Is the binding site defined from a structure, pocket detector, or known ligand?
• Are clashes, bond geometry, chirality, and protein-ligand contacts validated?
• Does the model condition on the full protein, local pocket, surface, or residue graph?
• Are pose quality and binding affinity evaluated as separate targets?
• Are ligand scaffolds and protein families both controlled in the split?
• Is the complex experimental, docked, predicted, generated, or used only as a reference?
• Does the label come from the complex geometry, an assay, or a curated benchmark?
• Is the protein structure apo, holo, predicted, or transferred from a template?
• Are water, metals, cofactors, and protonation states included or removed by protocol?
• Are invalid poses counted as failures rather than filtered after scoring?

Related

• Entity relation map
• Protein
• Pocket
• Ligand
• Target-assay-label contract
• Protein-ligand interaction
• Interaction fingerprint
• Pose generation
• Pose quality
• Binding affinity
• Docking workflow
• Protein family split
• Scaffold split
• Protein-ligand docking