Peptide-MHC structure prediction by SE(3)-equivariant diffusion, sampling 10 Cα conformations in under 6 seconds at 0.45 Å best-of-10 RMSD.
No providers recorded yet. Browse all providers
MHC-Diff predicts the three-dimensional conformation a peptide adopts inside the binding groove of an MHC molecule. Major histocompatibility complex proteins display intracellular peptides on the cell surface, where T-cell receptors scan them to distinguish healthy cells from cancerous or infected ones. Knowing the geometry of that displayed peptide matters for neoantigen vaccine design and T-cell therapy, but the problem is hard: HLA loci carry over 40,000 known alleles, each with distinct length and sequence preferences, and the space of tumour-derived peptides is combinatorially large.
The model was developed by David Frühbuß with the Xue lab at Radboud University Medical Center and collaborators at the University of Amsterdam, the Max Planck Institute of Biochemistry, TU Munich, and MIT, and posted to bioRxiv in May 2025. It differs from earlier structure-based approaches in two ways. First, it is generative: rather than emitting one static pose, it samples a distribution of plausible peptide conformations, which reflects the real conformational flexibility of peptide–MHC binding and matters for downstream TCR-recognition modelling. Second, it is fast enough for screening — the design targets throughput rather than the per-complex accuracy ceiling that motivates AlphaFold-based pipelines.
MHC-Diff is distinct from the other peptide-MHC entries in this catalog. MHC-Fine is an AlphaFold fine-tune trained through OpenFold, alphafold_finetune is the earlier anchor-free AlphaFold adaptation from Motmaen et al. that MHC-Diff benchmarks against, and mhcSFM is a contrastive retrieval model rather than a structure generator.
Training uses two curated datasets. The 8K set holds 7,726 peptide–MHC structures on the single allele HLA-A*02:01 with 9-mer peptides — 202 experimental X-ray structures from the PDB plus 7,524 PANDORA-modelled complexes — split by Gibbs-sampling peptide clusters under leave-one-cluster-out cross-validation. The 100K set spans roughly 100,000 PANDORA structures plus 1,000 IMGT X-ray structures across diverse alleles and peptide lengths, split by hierarchical clustering over MHC alleles so that every test allele is unseen; the held-out test set contains 4,007 samples including 72 experimental structures. Only the MHC binding-groove residues are retained as conditioning context, and all complexes are pre-aligned to a reference pocket frame.
On the 8K benchmark, MHC-Diff reaches a best-of-10 Cα-RMSD of 0.45 Å, and its worst-case error stays below 2 Å where the AlphaFold2 fine-tune produces a 9.60 Å outlier. On the harder 100K benchmark with entirely unseen alleles it reaches 0.61 Å best-of-10 and 1.50 Å mean — still below the 0.84 Å that APE-Gen 2.0 achieves on the easier single-allele set. Generating 10 conformations takes 5.6 seconds, about 0.56 seconds per structure. Training consumed roughly 800 A100 GPU hours.
The intended use is high-throughput structural screening in cancer immunotherapy: modelling thousands of candidate tumour neoantigens against a patient's HLA type to prioritize peptides for personalized vaccines or T-cell therapy, and supplying structural input to binding-affinity and immunogenicity predictors that sequence-only tools such as NetMHCpan cannot provide. Sampling an ensemble rather than one pose is directly useful for TCR-engagement modelling, where the range of accessible peptide conformations is the quantity of interest.
MHC-Diff shows that a domain-specialized generative model can beat general structure predictors on their own metric in a narrow setting while running orders of magnitude faster, which is the trade that makes population-scale immunopeptidome screening tractable. Two limitations bound its use. The model outputs only Cα positions, so full backbone and side-chain reconstruction requires an external tool such as PDBfixer or MODELLER, adding roughly a second per structure. Around 12% of sampled candidates diverge badly — typically placing the peptide outside the groove — though these are easily filtered by RMSD or pocket occupancy. Coverage is limited to MHC class I; class II is named as future work. The preprint has not yet been peer reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.