Joint CDR sequence–structure design
Co-design CDR sequence и геометрии.
Generative taskБиологическая и инженерная цель
Согласовать identity residues CDR с loop geometry и, часто, antigen interface.
Возможные исходные данные
- framework
- antigen complex context
- CDR masks
Что именно генерируется
- CDR sequence + backbone/full-atom geometry
Классы генераторов
- joint-sequence-structure
- target-conditioned-binder
Какие архитектурные подходы используются
| Стратегия | Что генерирует | Представление | Примеры | Ограничение | Sampling | Novel topology | Integrated scoring | Complexity |
|---|---|---|---|---|---|---|---|---|
| Diffusion co-design | Sequence и CDR geometry | Equivariant 3D representation | DiffAb | Высокая стоимость sampling; у DiffAb нет prospective wet-lab | Joint denoising of type/position/orientation | В пределах CDR | Energy/structural metrics | Denoising steps × residues |
| Full-atom geometric generation | Sequence и атомные координаты | Atom/residue graph | dyMEAN, AbDiffuser | Сложная подготовка complex input | Iterative equivariant updates or diffusion | Ограничена antibody format | DockQ/ddG or physics priors | Atoms × iterations |
| Flow matching | Sequence–structure pair | Continuous geometric state | FlowDesign | Ограниченная независимая validation; public code NR | Flow-matching trajectory | В пределах CDR | Structure/energy metrics | Integration steps |
Architecture-to-workflow consequences
| Архитектурное решение | Практическое следствие |
|---|---|
| Requires complex structural context | Плохой input complex ломает generation. |
| Energy metrics in evaluation | Не равны measured affinity. |
Compute и deployment
- GPU required.
- Archived DiffAb repo ограничивает воспроизводимость.
Доступные инструменты
| Инструмент | Generated object | Публикация | Evidence |
|---|---|---|---|
| AbDiffuser | Full-atom antibody sequence + structure | Preprint | 3 prospective binding + expression validation |
| DiffAb | CDR sequence + backbone/geometry; interface proposal | NeurIPS 2022 main track (A*) | 1–2 in silico / retrospective |
| dyMEAN | Full-atom antibody sequence/structure and docked complex | ICML 2023 main track (A*) | 1–2 in silico / retrospective |
| FlowDesign | CDR sequence and structure | Peer-reviewed | 3–4 prospective binding/neutralization validation |
| RFantibody | CDR backbones, CDR sequences, antibody–target complex proposals | Peer-reviewed | 3–4 prospective binding and functional validation |
Типовые workflows
Evidence summary
DiffAb/dyMEAN — in silico; FlowDesign/AbDiffuser — узкий prospective context, с оговорками.
Что нельзя сравнивать напрямую
- Retrospective co-design ≠ prospective binder campaign.
Как выбирать
- Методологический co-design — DiffAb.
- Нужен открытый de novo antibody pipeline — RFantibody, а не DiffAb.