Antibody foundation model generating paired sequences with genetic and developability metadata for labelling, humanisation, and library design.
No providers recorded yet. Browse all providers
A therapeutic antibody has to satisfy many constraints at once — it must bind its target, look human enough to avoid an anti-drug antibody response, and carry surface properties compatible with manufacturing. Computational pipelines usually tackle these one at a time, chaining a humanisation tool to a developability filter to a germline annotator, each a separate model with its own training data. The sequential structure creates a real failure mode: a mutation introduced to fix hydrophobicity can undo a humanising substitution made two steps earlier.
AbBFN2, from InstaDeep, collapses that pipeline into one model by learning a single joint distribution over antibody sequence and its annotations. It is built on the Bayesian Flow Network paradigm, which differs from both autoregressive language models and diffusion models in what it operates on: rather than corrupting and denoising the data, a BFN maintains and refines the parameters of a distribution over the data, updating them through Bayesian inference as progressively less noisy observations arrive. Because those parameters live in a continuous space whatever the underlying variable, one network jointly models amino-acid tokens, categorical germline gene labels, and real-valued biophysical scores without modality-specific heads.
The practical payoff is conditional flexibility. AbBFN2 always generates all 45 of its variables at inference; fixing any subset as context defines a task. Fix the sequence and read off the annotations, and it is a labeller. Fix the CDRs and the species label, and it is a humanisation tool. Fix a germline lineage plus a developability target and generate freely, and it is a library designer. None of these require fine-tuning, which distinguishes it from antibody language models such as AbLang and p-IgGen that are typically adapted per task.
AbBFN2 has 800M parameters and was trained on 2,031,524 paired heavy/light sequences from the Observed Antibody Space, filtered for productive, complete variable domains and clustered by CDR3 identity plus 95% whole-chain similarity before a 99/0.5/0.5 train/test/hold-out split. Structures were modelled with ABodyBuilder2 to derive TAP developability metrics; germline annotations came from IgBLASTn on the nucleotide sequences. Training ran for 30,000 steps at batch size 4096, with clusters sampled in proportion to the square root of their size.
On annotation, it beats the alignment tools ANARCI and IgBLASTp on every task tested, reaching Pearson's R of 0.97 for heavy and light V-gene identity and F1 of 1.00 for V-gene family, species, and light-chain locus; it also predicts the heavy D-gene, which amino-acid level tools do not attempt. Inpainting VH–VL interface residues recovers 75.7% of the original amino acids with no significant difference in Rosetta interface ΔG. Its human-class logits correlate with observed clinical anti-drug antibody rates at R = −0.52, matching p-IgGen's −0.53. It humanised all 25 non-human clinical-stage precursors tested, using 46.8 ± 11.7 mutations against 45.1 ± 9.7 introduced experimentally, at a mean CDR RMSD of 0.69 Å. A conditioned library of 2,500 VRC01-like anti-HIV antibodies yielded 1,715 sequences meeting every requested genetic, loop-length, positional, and developability constraint — a combination held by only 21 of the 2M training sequences.
The model suits antibody engineering groups running lead-optimisation campaigns, where the usual sequence of humanisation, liability removal, and germline analysis can instead be posed as a single conditional generation. Its library mode is most valuable for target classes whose desired properties are too rare in natural repertoires to support fine-tuning a task-specific generator. Inference code and a hosted web application are both available.
AbBFN2 demonstrates that Bayesian Flow Networks handle genuinely heterogeneous biological data — sequence, categorical genetics, continuous biophysics — inside one generative model, and that multi-objective antibody design can be expressed as conditioning rather than as a pipeline. Two constraints bound its use: the work remains a preprint awaiting peer review, and while the inference code is Apache-2.0, the model weights and everything generated with them are licensed for non-commercial purposes only. All validation is computational, resting on predicted structures and in-silico developability proxies rather than expressed and measured antibodies.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.