De novo backbone → inverse folding
Практический сценарий
Нужна новая topology или motif scaffold, затем sequence.
Вход
- length or contig/motif
- optional target/hotspots
Pipeline diagram
Backbone generation
→Sequence design
→Self-consistency
→Experimental tests
Mermaid diagram
flowchart TD
A[Motif or length or target] --> B[3D diffusion generator]
B --> C[Backbone proposals]
C --> D[Inverse-folding model]
D --> E[Sequence-structure candidates]
E --> F[Independent folding]
F --> G[Experimental validation]Стадии pipeline
| Стадия | Архитектурный класс | Что происходит | Генерация или оценка | Generator | Промежуточный объект | Альтернативы |
|---|---|---|---|---|---|---|
| Backbone generation | SE(3)-equivariant diffusion | Генерируются backbone coordinates | generation | rfdiffusion | Backbones | FrameDiff, RFdiffusion All-Atom (RFdiffusionAA) |
| Sequence design | Graph inverse folding | Подбирается sequence | generation | proteinmpnn | Sequence–backbone pairs | ESM-IF1 |
| Self-consistency | Structure predictor | AF2/RF проверяет designability | evaluation | NR | Filtered designs | AlphaFold2 / ColabFold |
| Experimental tests | NR | Binding/function/structure | experiment | NR | Hits | NR |
Hardware и runtime
GPU practically required for diffusion. Runtime boundary: generation vs AF2 filter must be separated. Exact hours NR.
Evidence level
RFdiffusion: prospective general protein evidence. Vanilla model fails on antibody loops.
Точки отказа
- Hotspot mis-specification.
- Using vanilla RFdiffusion for antibodies.
- Treating pLDDT as affinity.
Критерии решения
- Antibody? Переходите на RFantibody.
- Ligand atoms? RFdiffusionAA + LigandMPNN.