Scoring Function

A scoring function assigns a value to a candidate structure, pose, or interaction. In docking, it often ranks protein-ligand poses or estimates binding-related quality.

A generic scoring function is:

$$s=S_{\theta }(P,L,X)$$

where $P$ is a protein or pocket, $L$ is a ligand, and $X$ is a candidate pose or complex geometry. Lower or higher scores may be better depending on the convention.

Uses

• Score candidate poses produced by pose generation.
• Rank candidate poses from protein-ligand docking.
• Estimate or proxy binding affinity.
• Filter physically implausible structures.
• Compare model outputs under a shared objective.

Score Types

Type	Scores	Typical Evidence
pose score	candidate pose geometry $X$	RMSD, pose plausibility, interaction recovery
affinity score	protein-ligand binding strength	correlation, RMSE/MAE, assay-aware split
ranking score	candidate list or virtual screening output	enrichment, AUROC, PR-AUC, top-k recovery
validity score	physical or chemical plausibility	steric clashes, bond lengths, valence, PoseBusters-style checks
learned energy or potential	geometry-dependent compatibility	calibration, OOD behavior, ablation against simple baselines

The score type must be stated before interpreting a result. A pose score can help choose a geometry but does not automatically prove affinity prediction.

Calibration and Convention

Record the score convention:

$$\mathrm{rank}(X_i)={\mathrm{argsort}}_i\left(\sigma S_{\theta }(P,L,X_i)\right),\sigma \in \{-1,+1\}$$

where $\sigma$ encodes whether lower or higher scores are better. Mixing conventions across tools, learned models, and benchmark scripts can invert rankings.

If the score is interpreted as probability or energy, record the calibration statement:

$$p_{\theta }(y=1\mid P,L,X)=\mathrm{calibrate}(S_{\theta }(P,L,X))$$

Without calibration evidence, a score is usually only ordinal within the evaluated protocol.

Failure Modes

• Good score with bad geometry.
• Dataset bias that rewards shortcuts.
• Poor calibration outside the training distribution.
• Confusion between pose quality and binding affinity.
• Score convention mismatch, where lower-is-better and higher-is-better are mixed.
• Scoring candidates produced by one generator and claiming generator-independent quality.
• Using ligand-defined pockets or known poses that are unavailable in deployment.
• Reporting only the best-of-many score without accounting for search budget.

Task Boundary

A scoring function is not a pose generator:

$$S_{\theta }(P,L,X)\ne p_{\theta }(X\mid P,L)$$

The first evaluates a given pose $X$; the second samples or searches for candidate poses. A workflow may use both, but evaluation should say which component is being measured.

Checks

• Is the score used for pose ranking, affinity prediction, virtual screening, filtering, or sample selection?
• Is lower or higher better, and is that convention consistent across all reported metrics?
• Is the score calibrated, or only useful for ranking within one candidate set?
• Are candidates generated by the same method for all comparisons?
• Does the benchmark separate pose quality, affinity, enrichment, and physical validity?
• Is the search budget fixed before comparing best-scoring candidates?

Related

• Structure-based drug discovery
• Pose generation
• Pose quality
• Binding affinity
• Virtual screening
• Metric selection
• Objective-metric alignment
• Structure-based modeling
• PoseBusters
• Equivariant GNN
• LLM Wiki