bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / protein / igcraft
Protein
University of CambridgeReleased March 2025

IgCraft

Antibody sequence generation model that samples paired human VH/VL chains, covering inpainting, inverse folding, and CDR grafting in one network.

The short version

  • —Humanizes non-human antibodies by grafting their CDRs into generated human frameworks
  • —A single checkpoint covers unconditional sampling, inpainting, inverse folding, and grafting
  • —Denoises heavy and light chains jointly, so any subset of positions can be held fixed
  • —Conditions on antibody and epitope backbone geometry whenever a structure is available
88Openness6Citations
0HF downloads
22GitHub stars
MITLicense

Where to run it

No providers recorded yet. Browse all providers

Therapeutic antibody engineering is fragmented into separate computational tasks, each with its own specialist model: generating new paired sequences, filling in a masked region, designing sequence onto a known backbone, and grafting non-human complementarity-determining regions into a human framework. Running four tools with four sets of assumptions is a real drag on discovery pipelines. IgCraft, from the Sormanni lab at the University of Cambridge, collapses all four into one network that samples paired human heavy and light variable domains under whatever conditioning information is on hand.

What makes this possible is the choice of generative framework. IgCraft is built on a discrete-variable Bayesian Flow Network, which denoises a continuous field of per-token logits rather than the tokens themselves. Like a diffusion model, and unlike an autoregressive language model such as p-IgGen, it imposes no generation order and decouples the number of sampling steps from sequence length. That order-agnosticism is what lets a model trained only to produce full-length sequences perform arbitrary-region inpainting at inference time, without a task-specific head. The work was released as a preprint in March 2025 and presented at the GEM workshop at ICLR 2025, with code and the final checkpoint under an MIT license. A companion Zenodo deposit adds the intermediate post-pretraining, pre-fine-tuning weights along with the full 13.6 GB training corpus and the test set, all under CC BY 4.0.

#Key Features

  • Unified task coverage: Unconditional sampling, sequence inpainting, structure-based inverse folding, and CDR motif scaffolding all run from one checkpoint, selected by what is supplied as conditioning rather than by swapping models.
  • Paired-chain modeling: A two-track transformer processes VH and VL in separate stacks that exchange information through gated cross-attention interaction blocks, so heavy/light pairing statistics are learned rather than assumed.
  • Optional structural conditioning: A structure encoder built on geometric multi-head attention accepts antibody backbone and epitope coordinates when they exist and is simply omitted when they do not.
  • Humanness by construction: Generation is constrained to the space of human antibody sequences, which is why grafted frameworks score higher on humanness metrics than sequences from species-agnostic design tools.
  • Length-flexible generation: Per-region padding tokens are generated alongside amino acids, letting the model choose CDR and framework lengths under conditional design scenarios.

#Technical Details

IgCraft contains roughly 300 million trainable parameters. The backbone uses gated self-attention, rotary positional embeddings, pre-layer normalization, and SwiGLU transitions, with adaptive layer normalization in the cross-chain interaction blocks; the structure encoder adopts the geometric attention design from ESM3. Sampling solves the BFN's logit-space SDE with a second-order solver in 20 steps, and conditional sampling uses particle filtering with 32 particles. Training proceeds in three stages: per-chain pretraining on 118 million unpaired VH and 135 million unpaired VL sequences from the Observed Antibody Space, fine-tuning on 1.5 million paired sequences, then structure fine-tuning on roughly 30,000 predicted and 2,800 experimental paired structures with only the structure encoder updated. On a 2,000-sequence holdout, unconditional samples match native repertoire novelty and diversity statistics. Inpainting recovers 91.6% of H-CDR1 and 41.1% of H-CDR3 residues, and 96.7% across all framework regions. On 98 held-out structures, inverse-folded sequences reach 45.0% H-CDR3 recovery — below AntiFold at 58.1% — but IgCraft is the only method whose designs improve both AbNatiV humanness and CamSol solubility over the native sequences on both chains. Grafting mouse CDRs from 27 paired structures yields an OASis humanness of 77.9 against 74.6 for HuDiff and 47.3 for the parental mouse antibodies, with H-CDR3 loop RMSD of 2.04 Å under AlphaFold 3 refolding and 10 of 20 bound complexes correctly docked.

#Applications

The immediate use is antibody humanization: taking a high-affinity mouse or other non-human binder and generating a human framework that preserves the CDR conformations responsible for binding, with the option of feeding CDR backbone geometry to the model when a structure or prediction exists. Beyond that, the same checkpoint supports library design from unconditional sampling, affinity maturation and developability repair through region-wise inpainting, and sequence design onto a designed or predicted antibody backbone. Because humanness and solubility of the outputs are competitive with dedicated tools, it fits as a single step in a workflow that previously chained several.

#Impact

IgCraft is the first paired-antibody generative model to combine unconditional and conditional sampling with optional structural conditioning in one trained network, and it makes a concrete case that Bayesian Flow Networks are a practical alternative to both autoregressive and discrete diffusion generators for biological sequences. Its strongest result is in CDR grafting, where it exceeds a diffusion-based humanization model on both humanness metrics despite not being fine-tuned to optimize either. Task-specific models still lead on individual benchmarks: AbLang2, the successor to AbLang, edges it on CDR inpainting recovery, and AntiFold on inverse-folding recovery. All evaluation is computational, relying on predicted humanness, solubility, and structure scores rather than binding measurements, and the paper appeared as a workshop contribution rather than a full peer-reviewed publication.

At a glance

Parameters
300 Million
Released
March 2025
Category
Protein
License
MIT
Organization
University of Cambridge

Related models

  • IgGM

    Tencent AI for Life Science Lab / Chinese Academy of Sciences / University of Chinese Academy of Sciences / Fudan University / Sun Yat-sen University / Shanghai Jiao Tong University / Guangzhou Medical University

  • AbBFN2

    InstaDeep

  • IgLM

    GrayLab

  • p-IgGen

    Oxford Protein Informatics Group (OPIG) / AstraZeneca

  • AbFlowNet

    Bangladesh University of Engineering and Technology / University of California, Riverside / University of Southern California / Princeton University

  • Igloo

    Prescient Design / Genentech / Harvard University

  • DoggifAI

    University of Edinburgh / University of Gdańsk / University of Calgary

Links

GitHub RepositoryResearch PaperHuggingFace ModelDataset

Tags

antibodybayesian_flow_networkinverse_foldingmotif_scaffoldingtransformer

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.