dyMEAN

dyMEAN dyMEANOpt

ICML 2023 main track (A*) Direct antibody flag

Summary card

Поле	Значение
Год / version	2023 / NR
Last verified	2026-08-13
Что генерирует	Full-atom antibody sequence/structure and docked complex
Форматы	Fv-like antibody complexes
De novo или optimization	CDR generation/redesign
DOI	NR
Paper	https://proceedings.mlr.press/v202/kong23c.html
GitHub / repo	https://github.com/thunlp-mt/dymean
Service / docs	NR
Лицензия	MIT
Доступность / commercial-use	Code + environment; external scorers needed

Задача и generated object

Used in these tasks: Joint CDR sequence–structure design, Antigen-conditioned antibody design

Part of these workflows: Antigen-conditioned CDR co-design

Поле	Значение
Required input	Epitope and incomplete antibody sequence
Conditioning	Epitope, incomplete antibody, CDR design region
Fixed	NR
Editable	NR
Generated	Full-atom antibody sequence/structure and docked complex
Single mutations	Possible
Combinatorial mutations	Yes
Multi-objective	No explicit Pareto optimizer

Architecture and operating principle

Author-reported (matrix)

Поле	Значение
Base architecture	NR
Model type	4 + 5 Joint sequence–structure target-conditioned generator
Representation	NR
Generation unit	Full-atom antibody sequence/structure and docked complex
Conditioning	Epitope, incomplete antibody, CDR design region
Output	Full-atom antibody sequence/structure and docked complex
Sampling	NR
Internal scorer	External TM-score/DockQ/ddG tools in evaluation
Structure generation	Jointly models/docks complex

Input representation

Поле	Значение
Node features	NR
Edge / pair features	NR

Encoder и decoder

Поле	Значение
Encoder	NR
Decoder	NR

Latent or generative state

Conditioning mechanism

Epitope, incomplete antibody, CDR design region

Generation mechanism

E(3)-equivariant multi-channel graph network with iterative sequence/structure updates

Sampling: NR

Training objectives

Losses: NR

Mathematical details: NR

Integrated evaluation

External TM-score/DockQ/ddG tools in evaluation

Architectural limitations

Predicted affinity optimization is not measured improvement.
training-data scarcity and epitope representation limit generalization.

Uses these architecture components: Graph neural networks, SE(3)-equivariant networks, Joint sequence–structure generation

Generator versus evaluation

Component	Role	Architecture class	Integrated or external	Required or optional	Influences generation or post hoc	Runtime included or excluded
dyMEAN equivariant generator	primary_generator	E(3)-equivariant multi-channel graph network with iterative sequence/structure updates	integrated	required	generation	included
External TM-score/DockQ/ddG tools in evaluation	integrated_scorer	NR	integrated	optional	post_hoc	unknown
Jointly models/docks complex	integrated_complex_predictor	NR	integrated	optional	post_hoc	included
Independent complex prediction and experiments	mandatory_external_dependency	NR	external	required	post_hoc	excluded
Experimental binding/function assays	experimental_validation	NR	external	required	post_hoc	excluded

Training data и biases

Поле	Значение
Training data	NR
Preprocessing	NR
Negative examples	NR
Fine-tuning	NR
Label requirements	NR
Training hardware	NR
Training time	NR

Inference hardware и runtime

Поле	Значение
Candidate count	NR
Batch size	NR
Generation steps	NR
Recycles	NR
Size	NR
Reported runtime	NR
Runtime boundary	NR
CPU	NR
GPU	NR
VRAM	NR
RAM	NR
Disk	NR
Databases	NR

Experimental evidence

Поле	Значение
Evidence class	1–2 in silico / retrospective
Retrospective	No prospective wet-lab validation in primary paper
Prospective binding	NR
Prospective function	NR
Developability	NR
Tested design count	No prospective wet-lab validation in primary paper
Success criterion	Retrospective structure/design metrics only
Independent validation	NR
Summary	Retrospective structure/design metrics only

Limitations and failure modes

Predicted affinity optimization is not measured improvement.
training-data scarcity and epitope representation limit generalization.

Deployment: Research-only / retrospective

Reproducibility, license, commercial use

Поле	Значение
Repo	https://github.com/thunlp-mt/dymean
License	MIT
Commercial-use status	Code + environment; external scorers needed

Alternative tools: NR

Complementary: NR

Related: NR

Not directly comparable with

AntiFold, Efficient evolution with ESM-1v, ESM-IF1, LigandMPNN, ProteinMPNN, Structural evolution (ESM-IF1 complex)

Сравнение допустимо только внутри совместимых категорий. Runtime и quality AntiFold vs RFantibody нельзя читать как одну шкалу: разные starting points и generated objects.

Depends on these external modules

Primary paper и полный список источников

End-to-End Full-Atom Antibody Design

Source matrix: landscape xlsx, last verified 2026-08-13.

Детальный source-grounded разбор

1. Что именно делает инструмент

[MATRIX]Вход: Epitope and incomplete antibody sequence

[MATRIX]Generated object: Full-atom antibody sequence/structure and docked complex

[MATRIX]Фиксируется: NR. Изменяется: NR.

[MATRIX]De novo или parent-based: CDR generation/redesign.

2. Архитектура в одном абзаце

[MATRIX]E(3)-equivariant multi-channel graph network with iterative sequence/structure updates

4. Математическая модель

NRТочные уравнения, losses и sampling solver будут внесены после чтения paper/code. Не подставляйте родовой UCB/transformer/diffusion, если он не подтверждён для этого инструмента.

Поля NR: conditional probability, update equations, sampling algorithm

5. Физическое устройство состояния

[MATRIX]Seed generated state / output: Full-atom antibody sequence/structure and docked complex. Structure generation: Jointly models/docks complex.

6. Физико-химическая интерпретация

[INTERPRETATION]Координаты, frames, PLM embeddings или residue tokens не являются расчётом ΔG, электростатики или растворителя, пока это не сделано внешней energy function. Likelihood / pLDDT / docking score ≠ экспериментальный binding.

7. Training

[MATRIX]Данные: NR. Fine-tuning: NR. Negative examples: NR.

8. Inference

[MATRIX]Sampling: NR. Generation steps: NR. Runtime boundary: NR.

10. Почему, по мнению авторов, метод эффективен

[MATRIX]Author-reported summary (матрица): Retrospective structure/design metrics only

NRCausal support не классифицирован, пока не прочитаны ablation и experimental sections.

12. Ablation analysis

NRAblation в этом выпуске не разобрана.

13. What the model does not represent

Явный растворитель и ионная сила — обычно отсутствуют в генераторе.
Протонирование, гликозилирование, кофакторы — обычно отсутствуют, если не оговорено.
Конформационная энтропия и долгая динамика — не моделируются одним forward pass.
Экспрессия, агрегация, иммуногенность — не равны likelihood генератора.
Истинная binding free energy — не равна internal score.

14. Evidence and limitations

[MATRIX]Evidence class (матрица): 1–2 in silico / retrospective. Prospective binding: NR. Function: NR. Developability: NR.

Predicted affinity optimization is not measured improvement.
training-data scarcity and epitope representation limit generalization.

15. Reproducibility

NRРепозиторий: https://github.com/thunlp-mt/dymean. Лицензия: MIT. Commit: NR, пока страница не имеет отдельного analysis YAML.

Модульная схема

Модуль	Класс модели	Вход	Выход	Обучается или frozen	Generator или evaluator
dyMEAN equivariant generator	E(3)-equivariant multi-channel graph network with iterative sequence/structure updates	NR	NR	NR	primary_generator
External TM-score/DockQ/ddG tools in evaluation	NR	NR	NR	NR	integrated_scorer
Jointly models/docks complex	NR	NR	NR	NR	integrated_complex_predictor
Independent complex prediction and experiments	NR	NR	NR	NR	mandatory_external_dependency
Experimental binding/function assays	NR	NR	NR	NR	experimental_validation

Generator versus evaluator

Компонент	Роль	Integrated или external	Required или optional	Влияет на generation или только ранжирует после
dyMEAN equivariant generator	primary_generator	integrated	required	generation
External TM-score/DockQ/ddG tools in evaluation	integrated_scorer	integrated	optional	post_hoc
Jointly models/docks complex	integrated_complex_predictor	integrated	optional	post_hoc
Independent complex prediction and experiments	mandatory_external_dependency	external	required	post_hoc
Experimental binding/function assays	experimental_validation	external	required	post_hoc

Claim–evidence matrix

Утверждение	Предполагаемый механизм	Тип evidence	Датасет / эксперимент	Comparator	Метрика / результат	Источник	Оговорка
Retrospective structure/design metrics only	E(3)-equivariant multi-channel graph network with iterative sequence/structure updates	2. retrospective benchmark	No prospective wet-lab validation in primary paper	NR	Retrospective structure/design metrics only	[MATRIX] landscape matrix 2026-08-13	Не читать как полный разбор primary paper.