DiffAb

DiffAb

Summary card

Задача и generated object

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

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

Architecture and operating principle

Input representation

Полный complex context: antigen residues и antibody framework как observed nodes; CDR nodes зашумлены. Это не sequence-only prompt.

Encoder и decoder

Equivariant GNN/transformer denoiser operating on 3D residues (Author-reported). Pairwise geometric edges to antigen provide target conditioning.

Latent or generative state

Joint discrete–continuous noisy state: amino-acid category, Cα coordinate, orientation frame.

Conditioning mechanism

Coordinate clamping of non-designed residues; residue masks; geometric edges to epitope; no classifier-guidance affinity head in the primary generator.

Generation mechanism

1) Зашумляются type, Cα и orientation designed residues. 2) Forward: discrete corruption + Gaussian/SO(3) noise. 3) Denoiser предсказывает cleaner type/geometry. 4) Sampling steps — paper/code default, exact number NR here. 5) Framework/antigen фиксированы. 6) Sequence и structure обновляются совместно. 7) Кандидат — decoded CDR on the given framework.

Sampling: Diffusion trajectories over designed CDRs while clamping context

Training objectives

Combined diffusion losses. No experimental ddG labels required. Training on SAbDab-like complexes implies dataset bias.

Losses: Category CE for residue type; coordinate/orientation denoising losses

Mathematical details: NR

Integrated evaluation

Primary paper is retrospective/in silico. Energy scores are evaluation, not measured affinity. Repository archived/read-only.

Architectural limitations

Uses these architecture components: Joint sequence–structure diffusion, SE(3)-equivariant networks, Coordinate diffusion

Generator versus evaluation

Training data и biases

Inference hardware и runtime

Experimental evidence

Limitations and failure modes

• Requires complex structural context.
• no prospective evidence.
• archived repository.
• energy scores are not affinity.

Deployment: Research-only / retrospective

Reproducibility, license, commercial use

Related tools

Alternative tools: dyMEAN, FlowDesign

Complementary: AntiFold, IgLM

Related: dyMEAN, FlowDesign, AbDiffuser

Not directly comparable with

Depends on these external modules

• DockQ / TM-score
• Experimental binding/function assays
• AlphaFold3

Related glossary terms

Primary paper и полный список источников

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

Детальный source-grounded разбор

Недоступные или неполные источники

DiffAb — antigen-conditioned joint diffusion CDR residue type, Cα и ориентации. Framework и antigen зажаты. Optional HDOCK и PyRosetta — не нейросеть. Archived repository.

Происхождение источников

Проверенные файлы реализации:

• diffab/models/diffab.py
• diffab/modules/diffusion/dpm_full.py
• configs/train/codesign_single.yml
• configs/test/codesign_single.yml
• README.md

1. Что именно делает инструмент

[DOC]Вход: antibody–antigen PDB, heavy/light IDs, config mode. Выход: CDR sequence + backbone geometry на маске. Фиксируется: framework и antigen (в codesign). Modes: codesign single/multi CDR, abopt, fixbb (sequence only), strpred (structure only). Parent complex-based, не sequence-only de novo.

2. Архитектура в одном абзаце

[CODE]DiffusionAntibodyDesign: ResidueEmbedding + PairEmbedding + FullDPM. EpsilonNet (GAEncoder, 6 layers) предсказывает rotation update, position epsilon и softmax по 20 АК. T=100. Position normalized /10 Å. Sequence embedding 25 types (padding).

4. Математическая модель

[CODE]См. класс joint sequence–structure diffusion. Потери rot cosine, pos MSE-on-noise, seq KL posterior. Sampling: полный шум на generate mask, затем 100 denoise steps.

5. Физическое устройство состояния

[CODE]На каждом t: (s_t, p_t, v_t) для каждого residue. Context — истинные (или observed) s,p,v. Fitness не в state.

6. Физико-химическая интерпретация

[DOC]Интерфейс может отражаться в pair geometry. H-bonds/electrostatics — не explicit. Optional PyRosetta binding energy — external approximate energy, не learned ΔG.

7. Training

[CODE]SAbDab chothia structures + summary TSV. mask_single_cdr. 200k iterations. Точный N комплексов и split hashes — NR без paper/processed data. Fine-tuning from general protein diffusion — не заявлено в просмотренном config (train from DiffAb model type).

8. Inference

[DOC]100 samples default test config. GPU required (качественно). HDOCK для antigen-only. Ray для parallel relax. Runtime boundary: neural sampling ≠ docking ≠ Rosetta ≠ experiment.

10. Почему, по мнению авторов, метод эффективен

NRProposed mechanism (from title/README): antigen-specific co-design via equivariant diffusion so sequence and structure remain consistent. Author claim: design and optimization of antibodies with diffusion models (NeurIPS). Evidence this pass: in silico / retrospective per catalog; не prospective wet-lab в primary paper по YAML/README. Causal support: unclear without ablation tables from PDF — marked NR.

12. Ablation analysis

[CODE]Ablations NeurIPS (fixbb vs codesign vs strpred как режимы inference существуют в configs, но количественные таблицы paper не извлечены). Наличие отдельных train flags train_structure/train_sequence позволяет режимы, не доказывая causal gain на binding.

13. What the model does not represent

• Экспериментальный binding
• Явный растворитель, гликаны, протоны
• Full conformational ensemble
• Developability
• Unbound antigen без HDOCK/template

14. Evidence and limitations

[INTERPRETATION]Archived repo (2023 last commit env.yaml). Energy scores are evaluation. Не сравнивать hit rate с RFantibody prospective campaigns.

15. Reproducibility

[DOC]Apache-2.0, weights on Hugging Face luost26/DiffAb. CUDA 11.3 env.yaml. SAbDab download required. Commit c3e29666.

Физико-химический смысл

Модульная схема

Generator versus evaluator

Claim–evidence matrix

Различия между статьёй и реализацией

Схемы information flow

flowchart LR
  subgraph gen["Generator"]
    DPM["FullDPM T=100"]
  end
  subgraph opt["Optional external"]
    HD["HDOCK"]
    PR["PyRosetta"]
  end
  subgraph exp["Experiment"]
    W["Wet-lab — not in paper as prospective"]
  end
  HD --> DPM
  DPM --> PR
  PR -.-> W

Источники

1. [1] Luo et al. NeurIPS 2022 DiffAb · источник
2. [2] GitHub luost26/diffab c3e2966601bf8025025ab87717b31b08fdd4834e · источник

Нерешённые научные вопросы

• Quantitative tables from NeurIPS
• Exact SAbDab date snapshot leakage