Loading…
Antibody language model that reads sequence and backbone coordinates together, so a masked CDR can be recovered from either or both modalities.
No providers recorded yet. Browse all providers
Someone re-engineering a CDR loop usually holds two pieces of evidence at once: the rest of the antibody's sequence, and a backbone for the variable domain that has either been solved or predicted. Antibody models have historically forced a choice between them. Sequence models such as AbLang, AbLang-2 and AntiBERTy read residues and never see geometry; inverse-folding models such as AntiFold and ESM-IF1 read a backbone and discard whatever sequence is already known. In both directions, half the available information is thrown away before the model is asked anything.
IgBlend, described by AstraZeneca's Centre for AI in BioPharmaceuticals R&D in an October 2024 preprint by Cedric Malherbe and Talip Uçar, removes the choice by making both modalities maskable. A residue can be replaced by the unknown token, and so can the three backbone coordinates (C-alpha, N and C) belonging to that residue. Pretraining mixes three objectives over the same weights — predict masked residues from sequence alone, from sequence plus structure, and from structure alone with the sequence fully blanked — so one checkpoint answers all three questions rather than three specialised models answering one each.
The practical consequence is four usable inference modes: sequence-only completion, sequence with the corresponding coordinates masked, sequence with the full backbone retained as guidance, and pure inverse folding. The model is deliberately antibody-specific rather than a general protein language model, and it is distinct from p-IgGen, the paired-chain generative antibody model AstraZeneca co-developed with OPIG.
A structure encoder passes backbone coordinates through four GVP-GNN layers and two transformer encoder layers to produce a rotation- and translation-invariant 512-dimensional embedding per residue; a sequence encoder produces a matching 512-dimensional embedding from one-hot residues through two transformer blocks with SwiGLU activations. The two are concatenated into a 1024-dimensional vector per residue and passed through a four-block, 16-head multi-modal trunk, then an MLP and softmax over amino acid types. Training data comes from SAbDab, PLAbDab, OAS and INDI: roughly 204 million unique sequences and 4 million unique structures, of which only about 4,000 are experimentally determined and the remainder inferred. Datasets were clustered with MMseqs2 and re-weighted so chain types are sampled evenly, with held-out clusters below 0.8 sequence identity to training data. Training ran 125,000 steps on 8 A10G GPUs.
On sequence recovery, structure-guided IgBlend reaches 61.5% CDR3 accuracy on heavy chains, 74.7% on light chains and 49.5% on nanobodies, against 53.8%, 68.2% and 33.1% for the strongest sequence-only baseline in each case. In inverse-folding mode it recovers heavy CDR3 at 53.6% versus 36.3% for AntiFold and 20.3% for ESM-IF1, and 73.6% of its generated heavy-chain sequences refold within 1 Å RMSD of the input backbone. On trastuzumab H-CDR3 variants with FACS-measured dissociation constants, sequence-only log-likelihoods correlate near zero with binding, while structure-guided scoring reaches a Spearman correlation of about 0.23 — comparable to AntiFold.
The model suits antibody engineering workflows where a backbone is already available: affinity maturation against a known complex, humanisation, and re-designing loops on a therapeutic scaffold. Its nanobody coverage helps single-domain campaigns, where sequence-only models trained mainly on conventional repertoires degrade sharply. Because scoring is zero-shot, log-likelihoods can pre-rank generated libraries before expression and binding assays.
IgBlend is one of the clearer demonstrations that synthetic structures — predicted rather than solved — are enough to teach an antibody language model useful geometry, since fewer than one in a thousand of its training structures came from crystallography. No code, weights or training data have been released, which limits independent reproduction; the preprint has not been peer reviewed, though workshop versions appeared at NeurIPS 2024. The model handles unpaired single chains only, with paired heavy-light fine-tuning left to future work, and the authors note that its structure-guided modes trade sequence diversity for recovery accuracy. AstraZeneca's own DiffAbXL benchmark, which ranks antibody design models by how well their scores track measured binding, places structure-only IgBlend among its stronger entries and sequence-only IgBlend well below them — consistent with the paper's finding that the structural channel carries the affinity signal.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.