RFantibody

RFantibody

Peer-reviewed Prospective flag Direct antibody flag

Summary card

ПолеЗначение
Год / version2025/2026 / NR
Last verified2026-08-13
Что генерируетCDR backbones, CDR sequences, antibody–target complex proposals
ФорматыVHH/nanobody and scFv; conversion of selected scFvs to IgG tested
De novo или optimizationDe novo CDR-mediated binder on a chosen framework
DOI10.1038/s41586-025-09721-5
Paperhttps://doi.org/10.1038/s41586-025-09721-5
GitHub / repohttps://github.com/RosettaCommons/RFantibody
Service / docsNR
ЛицензияMIT; free for commercial use
Доступность / commercial-useCode + weights

Задача и generated object

Used in these tasks: De novo antibody design, De novo VHH / nanobody design, Antigen-conditioned antibody design, Генерация CDR backbone, Joint CDR sequence–structure design

Part of these workflows: Target-conditioned de novo antibody (RFantibody)

ПолеЗначение
Required inputTarget structure, epitope hotspots, fixed antibody framework, CDR lengths
ConditioningFixed framework, hotspots, CDR lengths, VHH/scFv format
FixedAntibody framework and target (except designed CDRs)
EditableCDR backbones and sequences
GeneratedCDR geometry, CDR sequences, complex proposals
Single mutationsNo
Combinatorial mutationsYes across CDRs
Multi-objectiveNo explicit multi-objective optimizer

Architecture and operating principle

Author-reported (matrix + Nature paper status)
ПолеЗначение
Base architectureRFdiffusion (Ab) + ProteinMPNN + RF2
Model typeMulti-stage generative workflow
RepresentationSE(3) frames, then residue graph
Generation unitCDR frames then amino-acid tokens
ConditioningTarget, hotspots, fixed framework, CDR lengths
OutputSequence–structure antibody–target proposals
SamplingDiffusion trajectories + MPNN sampling
Internal scorerRF2 confidence / self-consistency
Structure generationДа (CDRs/backbone)

Input representation

Target structure + epitope hotspots + fixed antibody framework + CDR length contigs. Framework coordinates are clamped; CDRs are generated.

ПолеЗначение
Node featuresNR
Edge / pair featuresNR

Encoder и decoder

Equivariant RoseTTAFold-derived trunk as diffusion denoiser. ProteinMPNN is a separate encoder/decoder on the resulting backbone. RF2 is not a generator.

ПолеЗначение
EncoderRoseTTAFold-derived equivariant denoiser (antibody-fine-tuned)
DecoderReverse diffusion over frames then MPNN sequence decoder

Latent or generative state

Continuous noisy 3D frames for designed CDR residues, then discrete sequence tokens.

Conditioning mechanism

Coordinate clamping of framework and target; hotspot residue constraints; contig specification of CDR lengths; VHH vs scFv format choice.

Generation mechanism

1) Зашумляются CDR frames. 2) Forward process — SE(3) noise as in RFdiffusion. 3) Denoiser предсказывает cleaner frames. 4) Число sampling steps — implementation/default, exact default NR here. 5) Framework и target фиксируются. 6) Sequence назначается MPNN после backbone. 7) RF2 фильтрует комплексы.

Sampling: Multiple diffusion trajectories; MPNN samples per backbone; RF2/AF3 ranking

Training objectives

Antibody-fine-tuned diffusion objective on structural data; MPNN CE; RF2 trained as structure/complex predictor. Do not collapse these losses into one 'affinity loss'.

Losses: Diffusion score-matching / frame denoising (RFdiffusion family); MPNN cross-entropy; RF2 structure losses are predictor losses

Mathematical details: NR

Integrated evaluation

RF2 is an integrated complex predictor/filter. AF3 is described as retrospective/recommended external filter. Neither is experimental affinity. Display/SPR remain mandatory for binding claims.

Architectural limitations

Архитектурная особенностьВозможное ограничение
Fixed frameworkНет global Fv redesign.
Diffusion with many stepsВысокая inference cost.
Hotspot-conditioned generationОшибка epitope mapping почти фатальна.
Integrated RF2 rankingВозможна miscalibration вне training distribution.
No negative antigen examples as primary objectiveСпецифичность не доказана генератором.

Uses these architecture components: Frame diffusion, Inverse folding, SE(3)-equivariant networks

Generator versus evaluation

ComponentRoleArchitecture classIntegrated or externalRequired or optionalInfluences generation or post hocRuntime included or excluded
Antibody RFdiffusion + ProteinMPNNprimary_generatorAntibody-fine-tuned RFdiffusion → ProteinMPNN CDR sequence → fine-tuned RoseTTAFold2 filterintegratedrequiredgenerationincluded
Fine-tuned RF2 self-consistency/confidence; AF3 retrospective filteringintegrated_scorerNRintegratedoptionalpost_hocunknown
Fine-tuned RoseTTAFold2 includedintegrated_structure_predictorNRintegratedoptionalpost_hocincluded
AF3 recommended retrospectively; experimental binding/epitope/function tests mandatorymandatory_external_dependencyNRexternalrequiredpost_hocexcluded
Experimental binding/function assaysexperimental_validationNRexternalrequiredpost_hocexcluded

Training data и biases

ПолеЗначение
Training dataAntibody-fine-tuned RFdiffusion/RF2 (exact corpora: NR here)
PreprocessingNR
Negative examplesNR as an explicit specificity-training set
Fine-tuningYes, from general RFdiffusion/RF2 toward antibody geometry
Label requirementsNR
Training hardwareNR
Training timeNR

Inference hardware и runtime

ПолеЗначение
Candidate countNR
Batch sizeNR
Generation stepsNR
RecyclesNR
SizeNR
Reported runtimeNR
Runtime boundaryMust separate diffusion, MPNN, RF2, optional AF3, and wet-lab
CPUNR
GPURequired (qualitative; hours/VRAM NR)
VRAMNR
RAMNR
DiskNR
DatabasesNo MSA database required for the diffusion stage (Author-reported workflow character); RF2 details NR

Experimental evidence

ПолеЗначение
Evidence class3–4 prospective binding and functional validation
RetrospectiveNR
Prospective bindingCorrect epitope and designed geometry for selected structures; TcdB neutralization after maturation; initial affinities can be modest
Prospective functionCorrect epitope and designed geometry for selected structures; TcdB neutralization after maturation; initial affinities can be modest
DevelopabilityNR
Tested design count95-design low-throughput or 9,000-design display campaigns per target; overall target success 0–2%; multiple confirmed VHH/scFv binders
Success criterionCorrect epitope and designed geometry for selected structures; TcdB neutralization after maturation; initial affinities can be modest
Independent validationNR
SummaryCorrect epitope and designed geometry for selected structures; TcdB neutralization after maturation; initial affinities can be modest

Limitations and failure modes

Deployment: Limited antibody pilot

Reproducibility, license, commercial use

ПолеЗначение
Repohttps://github.com/RosettaCommons/RFantibody
LicenseMIT; free for commercial use
Commercial-use statusCode + weights

Alternative tools: tFold System, BoltzGen

Complementary: ProteinMPNN, AntiFold

Related: RFdiffusion, ProteinMPNN, tFold System

Not directly comparable with

AlphaProteo, BindCraft, BoltzGen, Efficient evolution with ESM-1v, 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

Diffusion model, Inverse folding, Epitope / hotspot, Framework, CDR, pLDDT / ipTM / PAE

Primary paper и полный список источников

Atomically accurate de novo design of antibodies with RFantibody

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

Review queue

Детальный source-grounded разбор

partially verified Последняя сверка источников: 2026-08-27. Nature HTML 10.1038/s41586-025-09721-5 просматривался (pipeline, prospective VHH/scFv, modest Kd, OrthoRep). Полные Methods/SI и repo commit не зафиксированы (GitHub commits/main timeout). Preprint bioRxiv 10.1101/2024.03.14.585103 — 429. Не описывать RF2 как generator.

Недоступные или неполные источники

ИсточникПочему недоступенЧто использовано вместо
Nature Methods/SI completeHTML частично; SI не выгруженNarrative Results pipeline; catalog numbers from matrix cross-check
bioRxiv 10.1101/2024.03.14.585103HTTP 429Nature version
RosettaCommons/RFantibody commit SHAAPI timeoutURL репозитория без SHA

RFantibody — многостадийный workflow: antibody-fine-tuned RFdiffusion (CDR frames) → ProteinMPNN (sequence) → antibody-fine-tuned RoseTTAFold2 (filter). AF3 в исследовании — retrospective filter. Binding устанавливается assay/cryo-EM, не pLDDT.

Происхождение источников

МеткаИсточникСсылкаПримечание
[PAPER]Bennett, Watson et al. Nature 10.1038/s41586-025-09721-5https://doi.org/10.1038/s41586-025-09721-5NR
[DOC]https://github.com/RosettaCommons/RFantibody MIThttps://github.com/RosettaCommons/RFantibodyNR
ПолеЗначение
Репозиторийhttps://github.com/RosettaCommons/RFantibody
CommitNR (timeout)
Ветка / tagmain
ЛицензияMIT (каталог YAML)

Проверенные файлы реализации:

1. Что именно делает инструмент

[PAPER]Вход: структура мишени, epitope hotspots, фиксированный antibody framework, длины CDR / contig, формат VHH или scFv. Выход: CDR backbones + sequences + комплекс-предложения. Фиксируется: framework и target (кроме designed CDRs). De novo CDR-mediated binder, не parent affinity maturation AL.

2. Архитектура в одном абзаце

[PAPER]Не одна сеть. (1) SE(3) diffusion denoiser на базе RFdiffusion, fine-tuned на antibody geometry, генерирует CDR frames при clamped FW/target. (2) ProteinMPNN назначает CDR sequences на полученный backbone. (3) Fine-tuned RF2 оценивает комплекс (self-consistency/confidence) и отфильтровывает кандидатов. AF3 — дополнительный retrospective ranker в статье, не generator.

4. Математическая модель

NRТочные score-matching уравнения Ab-RFdiffusion в этом выпуске NR (Methods). Классово: reverse SE(3) на designed frames. MPNN: $p(s\mid x_{\mathrm{bb}})$. RF2: structure/complex prediction loss, не $p(K_d)$.

5. Физическое устройство состояния

[PAPER]Стадия 1: noisy/clean residue frames CDRs. Стадия 2: amino-acid tokens. Стадия 3: predicted complex coordinates + confidence maps. Experimental fitness — только после display/SPR/cryo-EM.

6. Физико-химическая интерпретация

[PAPER]Steric complementarity может отражаться в learned geometry + later RF2. H-bonds/electrostatics/solvation не explicit в diffusion. RF2/AF3 confidence — статистика структур, не ΔG. OrthoRep maturation — экспериментальная эволюция после слабых начальных Kd.

7. Training

NRAntibody fine-tune RFdiffusion и RF2; MPNN стандартный sequence recovery. Точные датасеты, фильтры, hardware — NR без Methods. Negative antigens как specificity objective — не описаны как главный loss.

8. Inference

[PAPER]Кампании статьи: порядка 95 designs low-throughput или ~9000 display per target (цифры каталога/Nature narrative; уточнять по paper figures). Hit rate overall target success 0–2% (каталог). GPU hours/VRAM NR. Отделять diffusion wall-time от MPNN, RF2 recycles, AF3, wet-lab.

10. Почему, по мнению авторов, метод эффективен

[PAPER]Mechanism: antibody-specific fine-tune diffusion на CDR; hotspot-conditioned epitope; MPNN для sequences; RF2 filter для self-consistent complexes; experimental display/SPR/cryo-EM.

[PAPER]Claim: atomically accurate de novo antibodies (VHH/scFv); некоторые нейтрализующие после maturation (TcdB). Initial affinities often modest (tens–hundreds nM).

[INTERPRETATION]Causal support: prospective binding/structure — да для selected designs. Ablation «зачем fine-tune» vs vanilla RFdiffusion — нужно SI; здесь не утверждать цифры. RF2 ranking не доказан как причина binding, только как filter перед экспериментом.

12. Ablation analysis

NRПолные ablation tables Methods/SI в этом проходе не извлечены (NR). Не заполнять вымышленными процентами.

13. What the model does not represent

14. Evidence and limitations

[PAPER]Prospective: binders confirmed; cryo-EM для selected; hit rates modest vs miniprotein binder pipelines (authors). Framework fixed. Target holo conformation критичен. Independent validation вне авторов — NR здесь.

15. Reproducibility

NRMIT, GitHub RosettaCommons/RFantibody, weights claimed in catalog. Commit SHA NR (timeout). Inference example не прогонялся в этом выпуске.

Физико-химический смысл

Признак или операция моделиФизико-химическая интерпретацияПрямо / косвенно / внешнеВажное ограничение
Hotspotsepitope residuesconditioninguser-defined, can be wrong
CDR framesloop backbonegeneratedside chains later
RF2 PAE/pLDDT-like confidencepredictor self-consistencyevaluator≠ affinity

Модульная схема

МодульКласс моделиВходВыходОбучается или frozenGenerator или evaluator
Ab-finetuned RFdiffusionSE(3) frame diffusiontarget, hotspots, fixed FW, CDR contigCDR backbone framesfine-tuned generatorgenerator
ProteinMPNNinverse foldingbackboneCDR sequencestypically frozen pretrained MPNNintegrated generator stage
Ab-finetuned RoseTTAFold2structure/complex predictorsequence+complex hypothesiscoordinates + confidencefine-tuned predictorevaluator / filter
AlphaFold3 (study)complex predictorsequencesmodels/confidenceexternalretrospective evaluator
Display / SPR / cryo-EM / OrthoRepexperimentdesignsbinding, structure, matured Kdn/aexperimental validation

Generator versus evaluator

КомпонентРольIntegrated или externalRequired или optionalВлияет на generation или только ранжирует после
RFdiffusion (Ab)backbone generatorintegratedrequiredgeneration
ProteinMPNNsequence generatorintegratedrequiredgeneration
RF2complex predictor / rankerintegratedused as filterpost_hoc (does not denoise frames)
AF3external complex predictorexternaloptional/retrospectivepost_hoc
Assaysexperimentalexternalrequired for binding claimsnot generation

Claim–evidence matrix

УтверждениеПредполагаемый механизмТип evidenceДатасет / экспериментComparatorМетрика / результатИсточникОговорка
De novo VHH/scFv binders to specified epitopes can be generatedAb-RFdiffusion + MPNN + filters + experiment4. prospective binding validationmultiple targets; 95 or ~9000 designs/campaign (catalog)undesigned / failed targets; miniprotein pipelines (qualitative)binding, cryo-EM epitope; hit rate 0–2% overall (catalog)[PAPER] Nature Results (HTML narrative)Methods numbers need SI; modest initial Kd
Designed geometry matches experimental structures for selected casesframe diffusion + filtering4. prospective binding validationcryo-EM subsetdesign modelstructural agreement (details SI NR)[PAPER] NatureSelected successes, not all trajectories
Neutralization after OrthoRep maturation (TcdB)experimental evolution on designed starting points5. prospective functional validationTcdB campaignparent designed cloneneutralization / nM affinities after maturation[PAPER] NatureMaturation is not the neural generator
RF2/AF3 confidence equals affinity—1. только architectural rationalen/an/afalse[INTERPRETATION] scope ruleConfidence ≠ Kd

Различия между статьёй и реализацией

Версия статьиCommit репозиторияCheckpoint / modelФайлРазличиеПрактическое следствие
Nature 10.1038/s41586-025-09721-5NRNR checkpoints this passNRРеализация не сверена построчно с MethodsInference hyperparameters брать из repo configs после clone, не из этой страницы

Схемы information flow

flowchart TD
  subgraph gen["Generative stages"]
    RF["RFdiffusion Ab — CDR frames"]
    MP["ProteinMPNN — sequence"]
  end
  subgraph eval["Integrated / external evaluators"]
    R2["RF2 filter"]
    AF["AF3 retrospective"]
  end
  subgraph exp["Experiment"]
    D["Display/SPR"]
    C["cryo-EM"]
    O["OrthoRep"]
  end
  RF --> MP --> R2 --> D
  R2 --> AF
  D --> C
  D --> O
Generator vs evaluator vs experiment. Оригинальная схема. Источник схемы: Nature 10.1038/s41586-025-09721-5

Источники

  1. [1] Bennett, Watson et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature. DOI 10.1038/s41586-025-09721-5 · DOI · источник
  2. [2] RosettaCommons/RFantibody · источник
  3. [3] Dauparas et al. ProteinMPNN. Science 2022. Inverse-folding stage. · DOI

Нерешённые научные вопросы