ImmuFold

An antibody's binding surface is six hypervariable loops, and the hardest of them, CDR H3, is assembled by V(D)J recombination and junctional insertion rather than inherited from an ancestor. It has no evolutionary relatives to align against, which is precisely the signal that multiple sequence alignment supplies to general-purpose structure predictors. The second constraint is throughput: a method that takes minutes per structure is fine for one antibody and unusable for a discovery campaign that needs to look at a hundred thousand of them.

ImmuFold, from the International Digital Economy Academy in Shenzhen and presented at IEEE BIBM 2024, addresses both constraints with the same design decision. In place of an alignment it uses ImmuBERT, a 650M-parameter antibody language model pretrained on hundreds of millions of natural antibody sequences, and feeds its residue embeddings to a structure network that predicts all-atom coordinates directly. There is no MSA search and no template lookup, so a prediction resolves in roughly one second. The paper reports higher-quality antibody structures than both IgFold and AlphaFold 2.

Emitting side chains as well as backbone atoms is the other deliberate choice. Antibody folding methods that stop at backbone coordinates leave the user to repack side chains with a separate tool before the structure can be docked against an antigen or scored for affinity; ImmuFold produces the complete heavy-atom model in one pass. It is unrelated to ImmuneFold, a similarly named immune structure predictor from a different group.

#Key Features

• Antibody language model instead of alignment: ImmuBERT is pretrained on hundreds of millions of natural antibody sequences, so the repertoire statistics that alignment cannot reach for a recombined loop arrive through the encoder instead.
• All-atom output: the structure network predicts main-chain and side-chain coordinates together, giving a model directly usable for docking and interface analysis.
• Second-scale inference: predictions take approximately one second, which is what makes structural triage of a large sequence library tractable rather than aspirational.
• CDR-focused accuracy: the complementarity-determining regions, the part of the molecule every downstream analysis depends on, are where the reported gains over prior methods concentrate.
• Encoder validated beyond folding: on an antibody binding task, ImmuBERT's representations outperform existing language models, which the authors treat as the underlying reason ImmuFold folds well.

#Technical Details

ImmuFold is a two-stage system. ImmuBERT, the 650M-parameter encoder, is a masked language model over natural antibody sequences; the structure prediction network consumes its embeddings and predicts all-atom coordinates end to end, without the iterative alignment processing that dominates the runtime of MSA-based predictors. The reported comparison covers IgFold — itself a language-model-plus-graph-network design that predicts backbone atoms in under 25 seconds — and AlphaFold 2, with ImmuFold generating higher-quality structures in about a second.

The work is a six-page conference paper (BIBM 2024, pages 3054–3059) with no preprint version, and neither code nor weights have been released. IDEA's public model card for AbLingua, the group's antibody language model family, states that ImmuFold remains with the institute; the only IDEA antibody encoder distributed publicly is the 600M AbLingua checkpoint on HuggingFace, released under Apache-2.0. AbLingua was subsequently scaled to 1.7 billion parameters trained on 1.4 billion sequences and published separately.

#Applications

The intended use is therapeutic antibody discovery at repertoire scale. One-second all-atom prediction supports structural triage of large candidate panels, structural annotation of sequenced repertoires, and generation of starting geometries for antigen docking, epitope analysis, developability assessment, and affinity models that require side-chain packing. Antibody engineering groups are the direct beneficiaries, since the sequences they care about are exactly the ones for which experimental structures are scarce.

#Impact

ImmuFold is part of the shift in antibody structure prediction away from alignment-based pipelines toward domain-specific language models paired with direct coordinate heads, and it pushes that line to all-atom output at second-scale latency. Its practical reach is limited by release status: without public code or weights, the reported advantage over IgFold and AlphaFold 2 cannot be reproduced or extended by other groups, and the model cannot be run outside IDEA. The work's more durable contribution to the public record is its evidence that a large antibody-specific encoder carries representations strong enough to drive both structure prediction and binding tasks — a claim IDEA's later, openly released AbLingua models make testable.