Target-conditioned miniprotein binder (BindCraft)
Практический сценарий
Есть target structure и hotspots; нужен non-antibody binder.
Вход
- target structure
- hotspot residues
- binder length
Pipeline diagram
AF2-gradient hallucination
→MPNN redesign
→AF2 + PyRosetta filters
→Binding assays
Mermaid diagram
flowchart TD
A[Target + hotspots] --> B[AF2 hallucination]
B --> C[MPNN redesign]
C --> D[AF2/Rosetta filters]
D --> E[Ranked miniprotein binders]
E --> F[Experimental binding]Стадии pipeline
| Стадия | Архитектурный класс | Что происходит | Генерация или оценка | Generator | Промежуточный объект | Альтернативы |
|---|---|---|---|---|---|---|
| AF2-gradient hallucination | Differentiable structure model | Оптимизируется binder sequence/complex | generation | bindcraft | Draft binders | AlphaProteo, BoltzGen |
| MPNN redesign | Inverse folding | Sequences переписываются под геометрию | generation | proteinmpnn | Redesigned sequences | NR |
| AF2 + PyRosetta filters | Integrated scorer | Confidence/interface/energy filtering | evaluation | NR | Ranked candidates | PyRosetta / Rosetta |
| Binding assays | NR | Измеряется binding/function | experiment | NR | Validated binders | NR |
Hardware и runtime
GPU required. PyRosetta commercial terms separate. Exact runtime NR.
Evidence level
3–4 prospective binding/function; nanomolar examples. Not antibody.
Точки отказа
- AF2 optimization bias.
- Hotspot choice.
- Interpreting AF2 confidence as Kd.
Критерии решения
- Нужен antibody format? Нет.
- Commercial Rosetta license?