Structure-Based Modeling
Structure-based modeling은 molecular structure를 first-class object로 다룹니다. Classical docking과 conformer workflow뿐 아니라 AI-based pose generation, scoring, structure-aware generation도 포함합니다. 핵심은 모델이 molecule score만 잘 내는지가 아니라 protein structure, ligand geometry, protein-ligand complex를 물리적으로 그럴듯한 시스템으로 다룰 수 있는가입니다.
이 wiki에서 이 영역은 먼저 molecular modeling anchor입니다. 가장 강한 claim이 learned representation, generation, scoring, evaluation에 있을 때 AI method가 들어옵니다.
Sequence-only input이 중심이면 Sequence-based modeling에서 시작하고, coordinate, pocket, pose, conformer, complex가 중심이면 이 페이지에서 시작합니다.
Structure-based modeling은 AI 여부로 나누지 않습니다. Classical docking, force-field minimization, learned scoring, diffusion pose generation은 모두 같은 object와 evaluation problem을 공유합니다. 차이는 search space, scoring rule, learned component, evidence입니다.
유용한 추상화는 아래와 같습니다.
여기서 는 protein 또는 pocket, 은 ligand, 는 initial/noisy geometry, 는 predicted pose 또는 structure, 는 score 또는 property입니다.
Working Map
- object를 정의합니다: protein, ligand, pocket, complex.
- candidate structure를 generate 또는 refine합니다.
- physical/chemical plausibility를 확인합니다.
- candidate를 rank 또는 score합니다.
- realistic split과 generalization assumption 아래에서 평가합니다.
Method Family
| Family | Core operation | Evidence needed |
|---|---|---|
| Docking/search | conformer and pose search, scoring | pose quality, enrichment, preparation protocol |
| Force-field refinement | energy minimization or dynamics | force field, time step, convergence, artifacts |
| Learned scoring | neural score or affinity predictor | split, baseline, calibration, leakage audit |
| Generative pose model | sample/refine coordinates | validity, diversity, pose plausibility, denominator |
| Structure-aware design | generate molecule/protein conditioned on structure | chemical validity, binding evidence, novelty, filters |
Structure Workflow Map
| Stage | Question | Common failure |
|---|---|---|
| Receptor preparation | 어떤 structure, chain, protonation, missing residue, cofactor인가? | template 또는 ligand leakage |
| Ligand preparation | stereo, tautomer, protonation, conformer policy는 무엇인가? | inconsistent chemical state |
| Pocket definition | known, predicted, ligand-defined, blind, grid-based 중 무엇인가? | unavailable information 사용 |
| Pose generation | search, diffusion, refinement, docking, sampling 중 무엇인가? | plausible해 보이지만 strained geometry |
| Scoring | pose score, affinity, ranking, enrichment, filter 중 무엇인가? | score meaning collapse |
| Evaluation | RMSD, clash, interaction, affinity, screening, utility 중 무엇인가? | metric이 claim과 맞지 않음 |
Core Questions
- 언제 sequence-only representation으로 충분하고, 언제 3D structure가 필요한가?
- task가 pose quality, binding affinity, molecular generation, virtual screening 중 무엇인가?
- error가 geometric, chemical, data-driven, evaluation artifact 중 어디에서 오는가?
- benchmark가 real generalization을 측정하는가, related structure memorization을 측정하는가?
- classical force field, minimization, simulation protocol이 method의 일부인가, diagnostic일 뿐인가?
Claim Separation
| Claim | Do not confuse with |
|---|---|
| pose prediction | binding affinity prediction |
| docking score | calibrated free energy |
| geometry validity | biological utility |
| high enrichment | good absolute affinity |
| generated molecule validity | synthesizability or target selectivity |
Topics
- Protein
- Ligand
- Protein-ligand complex
- Geometry for Structure Modeling
- Pocket definition contract
- Protein-ligand docking
- Structure-based drug discovery
- Pose generation
- Scoring function
- Pose quality
- Binding affinity
- Virtual screening
- Force field
- Energy minimization
- Molecular dynamics
- PoseBusters
Evaluation Anchors
- PoseBusters: pose plausibility check.
- Pose generation: search/sampling과 scoring 분리.
- Leakage: data와 benchmark audit.
- Scaffold split: ligand-side generalization.
- Protein family split: protein-side generalization.
Paper Route
Specific paper는 Structure-based modeling papers, Protein modeling papers, Generative model papers에 둡니다. 이 페이지는 reusable route인 object, geometry, scoring, split, evidence에 집중합니다.