ImmuneFold

The CDR3β loop of a T-cell receptor is built by V(D)J recombination and junctional insertion, not inherited from an ancestor. It has no evolutionary relatives to align against — precisely the signal alignment-based folding methods depend on — and antibody heavy-chain CDR H3 has the same problem. The obvious alternative, training a structure predictor on immune proteins alone, hits a second wall: only a few hundred TCR structures exist in the Protein Data Bank, orders of magnitude fewer than the antibody structures that already strain the specialist models built on them.

ImmuneFold answers both constraints with transfer learning. It starts from ESMFold, which reads structure out of a frozen protein language model and needs no alignment, and adapts it to immune proteins using low-rank adaptation (LoRA). Only two low-rank matrices per weight matrix are trained, so the fine-tune costs a fraction of the memory and time a full update would, and the learned change is merged back into the original weights afterwards. The work comes from the Institute of Computing Technology, Chinese Academy of Sciences with collaborators at the University of Chinese Academy of Sciences, the Central China Institute of Artificial Intelligence, and the University of Waterloo, and was posted as a preprint in November 2024.

Two checkpoints ship from the same recipe — one for TCRs, one shared between antibodies and nanobodies — alongside a protocol that reuses the folded complex for a task the model was never trained on: scoring whether a TCR binds a given epitope.

#Key Features

• LoRA confined to the folding trunk: The ESM-2 language model and the structure module stay frozen; adaptation touches only the Evoformer-like trunk that learns single and pair representations — where alignment-derived constraints would normally enter.
• Separate TCR and antibody checkpoints: The two receptor families differ enough in available training structures and chain organization to warrant one fine-tune each rather than a single shared adapter.
• Antigen context through recycling: When an antigen structure is supplied, its distogram is computed once and held fixed across recycling cycles while the receptor's distogram updates, letting the bound conformation inform the prediction.
• Zero-shot binding prediction: TCR-epitope affinity is estimated by folding the TCR-pMHC complex and computing a Rosetta interface energy, sidestepping the overfitting that supervised binding predictors suffer on scarce, biased binding data.
• No inference overhead: Because the adapters are merged into the base weights, prediction costs the same as ESMFold — roughly three seconds per structure on an A100.

#Technical Details

LoRA is applied to the linear layers inside the folding trunk's attention-with-pair-bias, outer-product-mean, transition and triangular-update modules, with rank 16 for single representations and 8 for pair representations. That leaves 2.7% as many trainable parameters as full fine-tuning, halves memory use, and cuts training from about 20 days to about 10 on eight A800 GPUs. Chains are separated by a relative position offset of 512 rather than ESM-2's poly-glycine linker, and an interface FAPE term is added to ESMFold's loss. The TCR model trains on 730 STCRDab complexes and the antibody model on 9,829 SAbDab structures, both split by deposition date and deduplicated against the test sets. On 27 held-out TCR complexes, CDR3β RMSD is 1.31 Å against 1.44 Å for TCRmodel2, 1.57 Å for AlphaFold-Multimer, 1.77 Å for ImmuneBuilder and 3.67 Å for ESMFold, with 1.12 Å on other CDRs, 0.71 Å on frameworks and a DockQ of 0.84 for the α/β pose. Antibody CDR H3 reaches 2.65 Å across 325 test antibodies. Adding the antigen structure improves antibody CDR H3 from 2.91 Å to 2.57 Å and nanobody CDR3 from 3.00 Å to 2.46 Å. The zero-shot binding protocol scores an AUROC of 0.69 across eight epitope subsets, against 0.55 for PanPep and 0.57 for TEIM.

#Applications

The immediate uses are structural: modeling an engineered TCR against a candidate neoantigen, resolving a nanobody CDR3 before affinity maturation, or building an antibody-antigen complex when the epitope is known. Because inference takes seconds rather than the minutes an alignment search demands, it also runs at screening scale: the authors folded all 32,703 non-redundant TCR-peptide pairs in VDJdb and released the structures, a thirty-fold expansion over the 937 complexes determined experimentally as of June 2024.

#Impact

The broader argument is methodological: parameter-efficient fine-tuning transfers cleanly to Evoformer-style structure predictors, not just sequence language models, putting specialization of large folding networks within reach of labs that cannot afford a full retrain. The honest limits are that predictions are single static structures, so the conformational changes TCRs undergo on binding are not captured; that antibody-antigen docking remains weak when the epitope is unknown, since the underlying language model saw only monomeric proteins; and that the binding-energy step depends on PyRosetta, free only for academic use. The work remains a preprint awaiting peer review.