Pose Generation

Pose generation produces candidate ligand geometries inside a protein pocket. It is the search or sampling part of docking, separate from scoring, affinity prediction, and virtual-screening ranking.

A pose generator can be written as a conditional distribution:

$$\hat{X}\sim p_{\theta }(X\mid P,L,C)$$

where $P$ is the protein or pocket, $L$ is the ligand, $C$ is optional context such as constraints or an initial geometry, and $\hat{X}$ is a generated ligand pose.

What It Must Specify

• Binding-site definition: known pocket, predicted pocket, blind docking, or constraint-based search.
• Ligand state: standardized molecule, tautomer, protonation, stereochemistry, and conformer policy.
• Protein state: receptor structure source, chains, missing residues, side-chain handling, and flexibility assumption.
• Coordinate frame: how the pocket and ligand are centered, rotated, cropped, or augmented.
• Sampling policy: number of candidates, diversity, refinement, and filtering.

Pose Variables

A ligand pose includes rigid and internal degrees of freedom:

$$X=T(R,t,\tau ;L)$$

where $R$ is rotation, $t$ is translation, $\tau$ are torsion or conformer variables, and $T$ maps ligand identity and pose variables to 3D coordinates.

Sampling Contract

Pose generation is a conditional sampling problem:

$$\{{\hat{X}}^{(k)}{\}}_{k=1}^K\sim {\pi }_{\theta }(\cdot \mid P,L,C)$$

where $K$ is the number of attempted poses. Evaluation should preserve the denominator:

$$\text{valid-pose rate}=\frac{\#\text{valid generated poses}}{\#\text{attempted poses}}$$

Reporting only the best retained pose hides invalid generation, failed refinement, and filtering cost.

Generator Families

Family	Pose Search Shape	Risk
classical docking search	heuristic/global-local search over pose variables	scoring and search failures are entangled
conformer + placement	generate conformers then align/place	conformer library bias
diffusion or flow	iterative coordinate/torsion denoising or transport	sampler budget and geometry validity
autoregressive torsion	sequential torsion or atom placement	order dependence and accumulated errors
refinement model	starts from initial pose and updates coordinates	depends on initial pose distribution

The generator family should be separated from the scorer. A better score can rescue poor generation, and a better generator can be hidden by a weak score.

Coordinate Frame Boundary

If the pocket is centered or aligned using the reference ligand, the task is no longer blind docking. A public note should state what frame is available at inference:

Frame Source	Deployment Meaning
receptor-only pocket center	usable for known pocket docking
predicted pocket center	depends on pocket predictor quality
ligand-defined box	can leak reference pose unless available from protocol
template ligand alignment	template-leakage risk
random or canonical frame	requires equivariant or augmentation-aware model

For coordinate-generative models, the output should transform consistently under rigid motion:

$$\hat{X}(RP+t,L)=R\hat{X}(P,L)+t$$

when the problem is rotation/translation equivariant.

Evaluation

Pose generation should be evaluated before scoring claims:

• Pose validity: bond lengths, chirality, clashes, ring geometry, and pocket overlap.
• Native-pose recovery when a valid reference pose exists.
• Diversity among generated poses.
• Robustness to receptor preparation and ligand state choices.
• Generalization under scaffold, protein-family, and template-leakage controls.

Pose Selection

If multiple poses are generated, the reported pose usually depends on a selection rule:

$$k^{\text{*}}=\arg \underset{k}{\max }S(P,L,{\hat{X}}^{(k)})$$

where $S$ can be a scoring function, confidence model, energy, or heuristic filter. Pose generation quality and pose selection quality should be evaluated separately when possible.

Failure Modes

• Generating a chemically invalid ligand geometry.
• Placing the ligand in a leaked ligand-defined frame unavailable at deployment.
• Producing diverse-looking poses that all violate the same pocket constraint.
• Treating a high scoring pose as physically plausible without validation.
• Reporting affinity or enrichment without separating pose generation quality.
• Counting only selected poses rather than all attempted poses.
• Using a pocket box, template, or initial pose unavailable in deployment.

Checks

• How many candidate poses are attempted, retained, refined, and scored?
• Is the coordinate frame receptor-only, predicted-pocket, template-based, or ligand-defined?
• Are ligand chemical state and receptor preparation fixed before generation?
• Is the generator evaluated separately from the selector/scorer?
• Are invalid, filtered, and failed samples included in the denominator?
• Are symmetry-aware atom mapping and pose validity checks used?

Related

• Protein-ligand docking
• Docking workflow
• Pose quality
• Pose RMSD
• Scoring function
• Receptor and ligand preparation
• Pocket definition contract
• Template leakage
• Coordinate frame
• Equivariance
• Flow matching
• Molecular generation
• PoseBusters