Docking Workflow

A docking workflow estimates plausible ligand poses in a protein binding site and scores or filters them for downstream use. A useful public note should separate pose generation, pose quality checks, scoring, and ranking claims.

A simplified workflow is:

$$(P,L)\to \{{\hat{X}}_1,\dots ,{\hat{X}}_K\}\to \{s_1,\dots ,s_K\}\to \mathrm{rank}$$

where $P$ is a protein or pocket, $L$ is a ligand, ${\hat{X}}_k$ are candidate poses, and $s_k$ are scores.

The first arrow is pose generation. The second arrow is scoring or filtering. Keeping those steps separate makes failures diagnosable.

Expanded as a contract:

$$\begin{array}{rl}(P_{\mathrm{raw}},L_{\mathrm{raw}})& \stackrel{{\varphi }_{\mathrm{prep}}}{\to }(P,L)\\ (P,L)& \stackrel{{\varphi }_{\mathrm{search}}}{\to }\{{\hat{X}}_k{\}}_{k=1}^K\\ (P,L,{\hat{X}}_k)& \stackrel{f_{\mathrm{score}}}{\to }{s}_k\\ \{(X_k,s_k){\}}_{k=1}^K& \stackrel{{\varphi }_{\mathrm{select}}}{\to }{\hat{X}}_{\mathrm{top}}\end{array}$$

Each arrow is a separate source of assumptions and failure.

Key Ideas

• Pose generation and scoring are different tasks.
• A good score is not enough if the generated pose is physically implausible.
• Docking boxes, protonation, conformer generation, and protein preparation can dominate results.
• Force-field minimization or pose relaxation should be reported as part of the method if it changes the output.
• Receptor and ligand preparation should be treated as part of the method, not as invisible preprocessing.
• Learned scoring functions should be evaluated separately from search and filtering heuristics.
• Public notes should describe generic workflow decisions, not private targets or unpublished results.

Workflow Variants

Variant	Main output	Typical evidence
pose prediction	top pose geometry	pose RMSD, PoseBusters-style validity
virtual screening	ranked compounds	enrichment, AUPRC, recall at budget
affinity prediction	binding score or $\Delta G$	regression/correlation and split control
rescoring	new ranking over fixed poses	pose source and score-only ablation
docking plus relaxation	refined pose	report raw and relaxed metrics separately

Do not compare workflows unless preparation, search budget, pose count, and failure handling are aligned.

Claim Separation

Claim	Required separation
”better search”	same scoring and filtering, different pose generation
”better scoring”	same generated poses, different score function
”better preparation”	same search/scoring, different preparation policy
”better end-to-end docking”	report each component and the final metric

This prevents a score function from receiving credit for better search settings, or a pose generator from receiving credit for a hidden filter.

Practical Checks

• What protein structure and binding site definition are used?
• How are ligand states, conformers, charges, and stereochemistry prepared?
• How many poses are generated and how is diversity handled?
• Are poses minimized or relaxed after generation, and are raw invalid poses counted?
• Is the workflow evaluating pose prediction, enrichment, affinity, or prioritization?
• Are invalid poses filtered before ranking metrics are reported?
• Are baselines and split protocols appropriate for the generalization claim?
• Is the comparison budget-matched: number of poses, search time, conformers, and receptor states?
• Are pose prediction, affinity prediction, and screening metrics reported separately?
• Are preparation failures included in the denominator?

Related

• Protein-ligand docking
• Pose generation
• Pose RMSD
• Receptor and ligand preparation
• Protein-ligand interaction
• Binding site
• Pose quality
• Scoring function
• Virtual screening
• Force field
• Energy minimization
• PoseBusters