Antibody heavy-chain Bayesian Flow Network fine-tuned from ProtBFN, generating VH sequences and inpainting CDR and framework regions zero-shot.
No providers recorded yet. Browse all providers
Antibody language models such as AntiBERTy and AbLang2 are trained as masked models: a fraction of residues is hidden and the network learns to restore them. That objective makes them good at filling small gaps, but the gaps they see in training are scattered and short. Ask one to redesign a whole framework — all four framework regions of a heavy chain at once, about three quarters of the variable domain — and the masking pattern falls outside anything it was trained on.
AbBFN, from InstaDeep, approaches the problem generatively. It is a Bayesian Flow Network that learns the unconditional joint distribution of antibody heavy-chain (VH) sequences, obtained by fine-tuning the general-protein model ProtBFN on the Observed Antibody Space (OAS). Because a BFN refines the parameters of a per-position categorical distribution rather than committing to tokens in a fixed order, any subset of positions can be clamped at sampling time and the rest resolved by particle filtering. Inpainting is therefore zero-shot: one checkpoint, trained only to generate whole chains, handles every masking pattern a user cares to ask for.
The model was introduced alongside ProtBFN in a September 2024 preprint and published in Nature Communications in April 2025, with Alex Graves — who introduced Bayesian Flow Networks — among the authors. It is the first antibody model in InstaDeep's BFN line; the 2025 successor AbBFN2 extends the idea to paired chains and metadata.
AbBFN inherits ProtBFN's 650M-parameter encoder-only transformer, laid out like ESM-2 650M — 33 layers, 20 heads over a 1280-dimensional embedding, rotary positional embeddings — with the token embedding replaced by a linear projection over the input distribution. Fine-tuning used the unpaired heavy-chain OAS download of 21 February 2024, filtered for productive, fully-numbered sequences and then cleared of anything similar to either test set, leaving 195M training examples out of 248M. Training ran 100,000 steps at batch size 8192 with a maximum sequence length of 256, about three days on 64 TPU v4 chips.
On 20,000 held-out OAS heavy chains, AbBFN reaches amino acid recovery of 96.2/96.5/94.2/96.9% on framework regions H1–H4 and 90.4/88.5/43.1% on CDR-H1/H2/H3, matching AntiBERTy and AbLang2 region by region despite those models having been trained on the full OAS including these sequences. With all four framework regions masked at once it recovers 95.6%, against 79.9% for AntiBERTy and 86.0% for AbLang2. On the 10-fold SAbDab benchmark it leads every method trained without test-set overlap, at 70.3/64.9/31.5% across the three CDRs; AbBFN+, produced by 1,000 adaptation steps on the nine training folds, reaches 77.1/72.6/39.7% and tops all methods on CDR-H1 and CDR-H2.
The natural use is antibody engineering where part of a chain is fixed and part is open:
resampling a CDR loop on a known scaffold, grafting a binding region into a different
framework, or humanising by regenerating the framework around retained CDRs. Because the
conditioning set is arbitrary, a campaign can mask whichever positions the assay implicates
without training a new model for each pattern. Inference code and weights for AbBFN ship in the
same repository and HuggingFace repo as ProtBFN, runnable on CPU, GPU or TPU through a provided
Docker image, with an inpaint.py entry point that takes a FASTA file and a region to resample.
AbBFN was the paper's evidence that a general protein BFN transfers to a specialised domain, and it gave the line a concrete engineering target that InstaDeep pursued into AbBFN2. Its limits are worth stating plainly. Coverage is heavy chains only, so paired-chain design is out of scope. All results are computational — amino acid recovery against natural sequences, with no expression or binding data. CDR-H3 recovery trails the antibody-specific baselines, which the authors attribute to that loop's V(D)J junctional diversity and non-templated nucleotide addition being poorly represented in the general-protein pretraining, compounded by the baselines' exposure to the test distribution. Inpainting at the paper's 1,024-particle setting is expensive. The released weights carry CC BY 4.0 while the inference code is CC BY-NC-SA 4.0, which bars commercial use of the released implementation.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.