Structure-Based Drug Discovery
Structure-based drug discovery는 3D molecular structure를 사용해 target-ligand recognition, pose quality, binding affinity, candidate prioritization을 해석합니다.
이 wiki에서 SBDD concept는 Structure-based modeling을 받쳐주는 층입니다. 구체적인 project나 thesis direction이 생기면 Research page가 여기로 다시 연결됩니다.
SBDD Contract
SBDD note는 protein-ligand system을 아래 계약으로 고정한 뒤 method와 claim을 읽습니다.
| Part | Meaning | Typical question |
|---|---|---|
| protein or receptor | structure source, chain, missing residues, cofactors, protonation? | |
| binding site or pocket | known, predicted, ligand-defined, blind, template-derived? | |
| ligand or candidate set | stereo, tautomer, charge, conformer policy, library source? | |
| pose or complex geometry | reference, docked, generated, minimized, noisy, refined? | |
| pose quality rule | RMSD, clash, strain, interaction, geometry validity? | |
| score or predictor | pose score, affinity, ranking score, filter, uncertainty? | |
| metric | pose success, enrichment, correlation, calibration, validity, utility? | |
| split and leakage boundary | scaffold, protein family, complex pair, template, time, assay source? |
The first reading question is not whether a method is AI-based. It is which part of the contract it changes.
Route Map
| Question | Start | Then Check |
|---|---|---|
| how are receptor and ligand prepared? | Receptor and ligand preparation, Docking workflow | chemical state, missing atoms, receptor state |
| how is the pocket defined? | Pocket definition contract | known, predicted, ligand-defined, blind, template-derived |
| how are poses generated? | Pose generation, Pose RMSD | atom mapping and symmetry |
| is a predicted pose plausible? | Pose quality, Interaction fingerprint | geometry, contacts, denominator |
| what score is being optimized or ranked? | Scoring function, Binding affinity | label semantics and assay context |
| is this a screening claim? | Virtual screening, Negative set, Ranking metrics | candidate pool and early enrichment |
| is the benchmark clean? | Protein-ligand split, Template leakage | protein family, scaffold, complex, assay/source |
Core Concepts
- Pose generation
- Pose RMSD
- Scoring function
- Docking workflow
- Pocket definition contract
- Receptor and ligand preparation
- Protein-ligand interaction
- Pose quality
- Interaction fingerprint
- Binding affinity
- Virtual screening
- Protein-ligand split
- Template leakage
Task Map
where is a protein or pocket and is a ligand.
Claim Separation
SBDD papers often use similar inputs but make different claims. Keep the claim narrow.
| Claim | Evidence needed | Do not confuse with |
|---|---|---|
| pose generation | reference pose, symmetry-aware RMSD, pose plausibility | affinity or enrichment |
| pose plausibility | clash, strain, interaction, chemical geometry | correct binding free energy |
| scoring | score separates better/worse poses or ligands | calibrated affinity |
| affinity prediction | assay-aware labels, split, uncertainty, baseline | docking score |
| virtual screening | candidate pool, negative set, early enrichment | absolute activity |
| structure-aware generation | generated validity, novelty, conditioning, filters | synthesizability or selectivity |
Denominators matter:
Filtering invalid poses or molecules before reporting ranking metrics changes the claim and must be stated.
Evaluation Denominators
| Evaluation | Numerator | Denominator |
|---|---|---|
| pose success | poses within threshold and passing plausibility checks | all generated or docked poses attempted |
| docking enrichment | actives recovered in early ranked set | full screened library or declared candidate pool |
| affinity correlation | examples with valid label and prediction | declared benchmark after preprocessing |
| generation validity | samples passing chemistry and geometry rules | raw generated samples before repair/filtering |
| interaction recovery | contacts or fingerprints matching reference | evaluated complexes with defined pocket and pose |
Public Note Template
| Field | Write |
|---|---|
| System | protein, pocket, ligand, complex, structure source |
| Preparation | receptor cleanup, ligand state, conformer policy, pocket rule |
| Method | docking, scoring, diffusion, refinement, screening, generation |
| Output | pose, score, rank, affinity, candidate set, generated molecule |
| Metric | RMSD, clash, strain, enrichment, correlation, calibration, validity |
| Split | scaffold, protein family, complex pair, template, assay/source, time |
| Leakage audit | ligand-defined pocket, template similarity, benchmark overlap |
| Public boundary | no private structures, internal target names, unpublished results, server paths |
Checks
- Is the task pose prediction, affinity prediction, enrichment, or molecule generation?
- Is pose generation evaluated separately from scoring?
- Are receptor and ligand inputs prepared consistently?
- Are pose quality and binding affinity evaluated separately?
- Is pose RMSD symmetry-corrected and separated from interaction or affinity claims?
- Does the benchmark split test scaffold, protein-family, temporal, or structure-level generalization?
- Is the protein-ligand split policy aligned with the claim?
- Could training data or template databases leak a close protein, ligand, or bound complex into evaluation?
- Are invalid generated structures filtered before ranking claims?