Peptide-MHC binding affinity and all-atom 3D structure in one attention network pass, reaching 1.19 Å median C-RMSD at 0.009 s per affinity call.
No providers recorded yet. Browse all providers
Working out which tumor peptides a patient's MHC molecules will display is the gating step in neoantigen discovery and T cell receptor therapy design. Sequence-only predictors answer quickly but say nothing about the shape the peptide adopts in the groove, which is what a TCR actually reads. Structure-based methods supply that shape and transfer better to under-characterised alleles, but modeling a complex has historically cost seconds to minutes, which rules out screening at proteome scale. SwiftMHC collapses the two into one network: a single forward pass yields both a binding affinity and an all-atom 3D model of the peptide-MHC complex.
SwiftMHC was developed by the Xue lab at Radboud University Medical Center with collaborators at the Netherlands eScience Center, the University of Amsterdam, and Wageningen University. It was posted as a preprint in January 2025 and published in Cell Reports Methods.
Several catalog entries attack pMHC structure from other directions: MHC-Diff samples peptide conformations with SE(3) diffusion, MHC-Fine fine-tunes AlphaFold weights through OpenFold, mhcSFM retrieves binders with a contrastive dual encoder, and alphafold_finetune reuses AlphaFold confidence as a binding proxy. SwiftMHC is a compact task-specific attention network trained from scratch on physics-generated structures rather than a repurposed general folding model.
The architecture has four modules. MHC residues are encoded by self invariant point attention over a one-hot sequence representation plus a pairwise residue proximity matrix; peptide residues are encoded by two rounds of self-attention with relative positional encoding; iterative cross-attention between the two then refines the peptide's frames; and a residue-wise MLP, summed over peptide positions, produces the affinity. Backbone rotations are carried as quaternions rather than 3×3 matrices, with an embedding dimension of 32 and up to 14 atoms per residue. All structures are superimposed on a reference MHC (PDB 3MRD) so frames start in a consistent orientation. Training pairs 7,726 IEDB affinity measurements for HLA-A*02:01 9-mers with PANDORA models selected by lowest molpdf score, and runs in two phases: frame-aligned point error, torsion angle, and affinity losses first, then a violation loss added to clean up bond geometry and clashes. Optimisation used Adam, batch size 16, a learning rate of 10⁻³, and gradient clipping at 0.5.
On binding affinity, SwiftMHC reaches a median AUC of 0.91, Pearson correlation of 0.80, and AUPR of 0.88, against 0.92, 0.81, and 0.89 for NetMHCpan 4.1 — comparable accuracy at 0.009 s per case in batch mode on a single A100 versus 0.081 s. On 202 held-out HLA-A*02:01 9-mer X-ray structures it achieves a median Cα-RMSD of 1.32 Å, improving to 1.19 Å after OpenMM energy minimisation, compared with 3.65 Å for MHCfold. An AlphaFold fine-tune scores 0.87 Å on the same set, but 105 of its 118 unique complexes had appeared in that model's training data.
The intended use is high-throughput triage of candidate tumor peptides: run a proteome's worth of 9-mers through the affinity head, then generate structures only for the survivors. Because the output is an all-atom complex, downstream analysis can go beyond a binding score — measuring which peptide atoms are exposed at the T-cell-facing surface, comparing a neoantigen to its self counterpart to anticipate cross-reactivity, or feeding the model into a TCR docking protocol. Groups building cancer vaccine or adoptive cell therapy pipelines are the natural users, and the Colab notebook lowers the barrier for immunologists without GPU infrastructure.
SwiftMHC's central claim is methodological: a small network trained on synthetic, physics-derived structures can match the structural accuracy of fine-tuned general-purpose folding models while running orders of magnitude faster, which suggests data scarcity in structural immunology is addressable by simulation rather than only by more crystallography. The practical scope is narrow — training and validation cover HLA-A*02:01 9-mers only, with no support for post-translational modifications — and affinity accuracy matches rather than exceeds NetMHCpan 4.1, so the gain there is speed and the accompanying structure. The stack also requires Linux and OpenFold, making deployment heavier than a pip-installable predictor.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.