MSA-free structure prediction for TCR-peptide-MHC complexes, pairing a protein-protein-interaction language model with a flexible docking module.
No providers recorded yet. Browse all providers
Recognition of a peptide presented by a major histocompatibility complex (pMHC) by an alpha-beta T cell receptor is the molecular event that licenses adaptive immunity, and it is the event that TCR-based cancer therapies and vaccines are engineered around. Modeling those complexes structurally has been hard for two reasons: the interaction surface is dominated by hypervariable CDR3 loops with little co-evolutionary signal, and the Protein Data Bank contains only a few hundred solved TCR-pMHC structures. General predictors such as AlphaFold-Multimer and AlphaFold 3 handle the problem, but at a per-complex cost dominated by multiple sequence alignment search — prohibitive when the goal is to model a repertoire of millions.
tFold-TCR, from the AI for Life Sciences Lab at Tencent with collaborators including Jamie Rossjohn's group, replaces the alignment step entirely. It builds on ESM-PPI-TCR, a 650M-parameter protein language model extended from ESM-2 by further pretraining on both monomers and multimers drawn from UniRef50, the PDB, protein-protein interaction and antibody databases, and a curated TCR collection. That model supplies intra-chain and inter-chain contact signal directly from sequence, which downstream structure modules turn into atomic coordinates.
The system is deliberately modular: separate feature generators predict the unliganded TCR and the unbound pMHC, and a representation-driven flexible docking module fuses their sequence representations, pair representations, and unbound structures into the complex — updating both partners' conformations along the way rather than treating them as rigid bodies.
Each feature-generation branch stacks ESM-PPI-TCR, an Evoformer-Single refinement module, and a structure module using invariant point attention for SE(3)-equivariant coordinate prediction. The TCR and pMHC branches share an architecture and initialization and differ only in training data. In the complex model these branches are frozen, and the flexible docking module — a 16-block Evoformer-Single stack followed by an 8-block structure module — predicts the assembled complex along with confidence estimates. Evaluation used temporal separation with a 1 January 2022 cutoff, on STCRDab-22-TCR (24 receptors), STCRDab-22-pMHC (18 pMHCs), STCRDab-22-TCR pMHC (18 complexes), and an in-house Rossjohn Lab Benchmark of 6 previously unpublished complexes.
On unliganded TCRs, tFold-TCR reaches framework-region backbone RMSDs of 0.566 Å (alpha) and 0.469 Å (beta), CDR-A3 RMSD of 1.665 Å and CDR-B3 of 1.352 Å, with overall RMSD 0.704 Å and GDT 0.956 — the most accurate of the methods compared. On complexes it scores 0.494 mean DockQ against 0.490 for AlphaFold 3 and 0.494 for AlphaFold-Multimer, and 0.474 versus 0.470 and 0.439 on the harder Rossjohn benchmark, with a success rate approaching 100% and no observed hallucination. Speed is where the margin is wide: roughly 30 times faster than AlphaFold 3 and over 2,000 times faster than AlphaFold-Multimer for complexes. Unbound pMHC prediction is the weak point, with DockQ of 0.908 against 0.927 and 0.926 for the MSA-based methods, driven by less accurate peptide side chains.
The intended use is immunology at repertoire scale: screening candidate TCRs against a target epitope, characterizing the structural basis of a known specificity, and building structural priors for TCR engineering in cell therapy and vaccine design. Because the model tolerates partial chain information and runs without alignments, it fits sequencing-driven workflows where thousands of clonotypes arrive at once and only paired CDR3 sequences are available. TCRStructDB itself provides a searchable structural resource for groups without the compute to fold their own repertoires.
tFold-TCR reframes TCR-pMHC modeling from a per-structure exercise into a repertoire-scale one, matching general MSA-based predictors on accuracy while removing the runtime barrier that made large-scale structural surveys impractical. Its accompanying database is the largest structural TCR resource assembled to date. The work is a preprint that has not been peer reviewed; accuracy on MHC II complexes lags MHC I because the open MHC II groove admits more binding-motif variability, side-chain contacts remain less reliable than backbone contacts, and the released code and weights carry a PolyForm Noncommercial license that rules out commercial use.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.