Docking
Docking estimates plausible ligand poses in a protein binding site and often ranks candidates for structure-based screening. In this site, docking is treated as a workflow with separate preparation, search, scoring, filtering, and evaluation steps.
where is the protein or pocket, is the ligand, is context such as pocket definition or constraints, and are candidate ligand poses.
A pose can be parameterized as translation, rotation, and internal torsions:
where places the ligand, orients it, and controls rotatable bonds. Docking searches over under a receptor or pocket context.
Workflow Map
| Step | Read |
|---|---|
| Define the object | Protein-ligand complex, Pocket, Ligand |
| Prepare inputs | Receptor and ligand preparation |
| Generate poses | Pose generation |
| Score candidates | Scoring function, Binding affinity |
| Check geometry | Pose quality, Pose RMSD, PoseBusters |
| Screen libraries | Virtual screening, Interaction fingerprint |
| Check generalization | Protein-ligand split, Template leakage |
Object Contract
Docking examples should state the object boundary before model claims.
| Field | Meaning |
|---|---|
| Protein or pocket | full receptor, cleaned receptor, predicted pocket, known binding site, or blind search region |
| Ligand identity | standardized molecule, tautomer/protonation/stereo state, conformer source |
| Pose target | crystallographic pose, curated reference pose, generated candidate pose, or no pose label |
| Search space | fixed pocket, flexible side chains, ligand flexibility, constraints, or blind docking |
| Scoring target | pose ranking, candidate ranking, affinity proxy, enrichment, or downstream selection |
| Split unit | ligand scaffold, protein family, complex pair, assay/source, or time |
Separation of Claims
Do not collapse these claims:
- pose generation: can the method place the ligand plausibly?
- pose quality: is the geometry chemically and physically plausible?
- scoring: does the method rank poses or candidates well?
- affinity prediction: does the method estimate binding-related labels?
- virtual screening: does the method enrich useful candidates early?
The same method can be strong for one claim and weak for another. A low pose RMSD does not automatically prove affinity prediction, and a high enrichment score does not automatically prove chemically valid poses.
Search and Scoring
Docking often combines a search distribution and a score:
The selected pose is usually:
or if the score convention is higher-is-better. This convention must be written explicitly.
Typical Metrics
| Claim | Typical Metric | Caveat |
|---|---|---|
| Pose generation | RMSD to reference pose | symmetry correction and atom mapping matter |
| Pose plausibility | geometry checks, steric clashes, bond validity, interaction sanity | a plausible pose can still rank poorly |
| Scoring / ranking | Spearman, Kendall, top-k success, enrichment | score calibration is separate from ranking |
| Affinity prediction | RMSE, MAE, Pearson/Spearman, calibration diagnostics | assay noise and label semantics dominate |
| Virtual screening | enrichment factor, BEDROC, ROC-AUC, PR-AUC | active/decoy construction can bias conclusions |
What Goes to Papers
Keep this page for reusable docking concepts. Put a page under Structure-based modeling papers when the note is about a specific docking method, benchmark, or empirical result. For example, PoseBusters belongs in Papers because it is a specific evaluation paper; pose plausibility checks belong here and in Pose quality.
Checks
- Is the pocket known, predicted, ligand-defined, or blind?
- Are ligand scaffold and protein family split units both considered?
- Are protonation, tautomer, stereochemistry, and conformer choices recorded?
- Is RMSD symmetry-corrected and separated from interaction quality?
- Is a docking score being treated as affinity without evidence?