Pose Quality

Pose quality measures whether a predicted ligand pose is geometrically and chemically plausible in a protein binding site. It should be evaluated separately from binding affinity and interpreted alongside Pose RMSD.

A common native-pose comparison is ligand RMSD:

$$\mathrm{RMSD}(X,\hat{X})=\sqrt{\frac{1}{N}\sum _{i=1}^N\Vert x_i-{\hat{x}}_i{\Vert }_2^2}$$

where $X$ is the reference ligand pose, $\hat{X}$ is the predicted pose, and $N$ is the number of matched ligand atoms.

Validity can be treated as a conjunction of checks:

$$\mathrm{valid}(\hat{X})=\prod _{k=1}^{K}1[c_k(\hat{X})=1]$$

Quality Axes

Pose quality is multi-axis. A single RMSD number is not enough.

Axis	Question	Typical Check
atom mapping	are equivalent atoms and symmetries handled?	symmetry-corrected RMSD
internal geometry	is the ligand chemically plausible?	bond lengths, angles, chirality, ring planarity
protein-ligand geometry	does the pose avoid impossible contacts?	steric clashes, contact distances, interactions
pocket consistency	is the pose in the intended binding site?	pocket overlap and interaction fingerprint
receptor state	is the receptor compatible with the pose?	apo/holo state, side-chain clashes, missing residues
preparation state	are protonation, tautomer, charge, and hydrogens consistent?	chemical state contract

RMSD Boundary

RMSD needs an alignment and mapping policy:

$${\mathrm{RMSD}}_{\pi ,R,t}=\sqrt{\frac{1}{N}\sum _{i=1}^N{\Vert R{x}_i+t-{\hat{x}}_{\pi (i)}\Vert }_2^2}$$

where $\pi$ maps equivalent atoms and $(R,t)$ defines the alignment. For ligand poses in a fixed receptor frame, receptor alignment and ligand atom symmetry should be stated.

Validity and Filtering

If invalid poses are filtered before scoring:

$${\mathrm{score}}_{\mathrm{kept}}\ne {\mathrm{score}}_{\mathrm{attempted}}$$

Report both attempted and kept poses when the claim is about a generation or docking pipeline. Otherwise a strong score may reflect the filter rather than the model.

Checks

• Is RMSD reported only after ligand identity and symmetry handling are correct?
• Are bond lengths, stereochemistry, ring planarity, and clashes checked?
• Are protein-ligand contacts plausible?
• Are pose quality, ranking, and affinity metrics reported separately?
• Are invalid generated poses excluded before using scores for downstream ranking claims?
• Are failed poses counted in the denominator?
• Is the receptor aligned, fixed, flexible, apo, holo, or predicted?

Metric Boundary

Pose quality asks whether a candidate geometry is plausible:

$$q=Q(P,L,\hat{X})$$

It does not by itself prove binding affinity, selectivity, or screening utility. A native-like pose can still bind weakly, and a strong binder can be missed if the receptor or ligand state is wrong.

Related

• PoseBusters
• Pose generation
• Pose RMSD
• Scoring function
• Binding affinity
• Protein-ligand docking