MHChron

Peptide-MHC binding is the gatekeeping step of computational immunogenicity assessment: only peptides forming a stable peptide-MHC (pMHC) complex reach the cell surface for T-cell surveillance. The field's long-standing weakness here is statistical rather than architectural. Binding data exist for fewer than 1% of the roughly 43,000 documented HLA allelic sequences, and what exists skews sharply toward 9-mers on class I and 15-mers on class II. Predictors trained on that distribution learn dataset-specific regularities that fail to transfer to under-represented alleles or atypical lengths, which is why published AUROCs of 94-99% routinely degrade under independent re-evaluation.

MHChron, from the University of Edinburgh School of Biological Sciences, attacks the problem from the data side. The authors assembled 18.8 million pMHC pairs from IEDB, the HLA Ligand Atlas and CEDAR, then deliberately subsampled to 42,012 complexes engineered for uniform coverage across peptide length (8-36 residues), allele (214 class I alpha-chain alleles plus 19 class II alpha-chain and 79 class II beta-chain alleles) and binding label — one dataset spanning both MHC classes rather than the usual class-specific split. On it they trained two models of matched capacity: MHChron-edgy, a structure-aware graph attention network, and MHChron-light, a graph-free set-attention model working from sequence embeddings alone.

The headline result is a negative one with practical consequences. MHChron-edgy stayed within roughly 2 percentage points of MHChron-light across most evaluation regimes, pulling ahead only when extrapolating to unseen allele clusters — so the gains came from dataset design rather than architectural complexity, and the cheaper sequence model is the one shipped to users.

#Key Features

• Unified across MHC classes: A single model handles both the closed class I groove and the open-ended class II groove, in place of the separate class-specific predictors the field usually maintains.
• Wide peptide-length coverage: Training peptides span 8 to 36 residues at uniform density, covering long class II ligands and atypical class I lengths that length-restricted tools exclude.
• Leakage-controlled evaluation: Results are reported under a stratified split, a leave-one-peptide-cluster-out split over Gibbs-clustered motifs, and a per-locus leave-one-allele-cluster-out split, with an accompanying leakage analysis.
• Whole-protein screening: The released package scans a full protein sequence for immunogenic peptides and produces interpretable visual output.
• Lightweight deployment: The user-facing model is a 41,666-parameter classifier over precomputed embeddings, installed from a single conda environment file.

#Technical Details

Both variants consume per-residue embeddings from ESM-2 — the 8M parameter esm2_t6_8M_UR50D checkpoint — computed separately for the peptide and each MHC chain after truncation to the peptide-binding domain. MHChron-edgy adds structure: TFold, an AlphaFold 2-based pMHC modelling pipeline, generates complexes that become residue graphs with 12 Å proximity edges weighted by inverse inter-residue distance, then two GATv2Conv layers (2 attention heads, 32 channels each) and global mean pooling, for 52,161 trainable parameters. MHChron-light drops the graph, masking padded positions across two self-attention blocks (2 heads, hidden size 32) and attention pooling, for 41,666 trainable parameters. Both end in a two-layer MLP trained with binary cross-entropy and assessed by five-fold cross-validation.

On the stratified split both models reached about 88% AUROC and 88-89% AUPRC against a 54.6% positive-label baseline. Under leave-one-allele-cluster-out evaluation, MHChron-edgy averaged 71% AUROC across class I loci and MHChron-light 65%; extrapolation to unseen class II alleles stayed near chance. Benchmarked on matched test subsets against NetMHCpan-4.2, NetMHC-IIpan-4.3, MHCfold and MHC-II3D — which improved AUPRC over their random baselines by 2 to 10 percentage points — both MHChron models gained 32 points. On 1,000 IEDB measurements deposited after model development, MHChron-light reached 81% AUROC and 93% AUPRC.

#Applications

The immediate uses are epitope discovery and neoantigen prioritisation, where candidates must be ranked against a patient's or a population's HLA background before synthesis. Coverage of both MHC classes and of long peptides also suits CD4+ helper-epitope work that class I-focused tools cannot address. The whole-protein scanner points the same model at a different question — deimmunising a therapeutic protein, gene-therapy vector or engineered enzyme by locating immunogenic hotspots in its sequence — putting it within reach of protein engineers who are not immunoinformatics specialists.

#Impact

MHChron's contribution is largely methodological: curating a compact, diversity-balanced training set and evaluating it under peptide- and allele-cluster holdouts buys more generalisation than adding structural inductive bias — a claim cutting against the direction structure-first pMHC work such as MHC-Fine has taken, and one that sequence-based peers like mhcSFM reach by another route. The splitting strategies and curated dataset, archived on Zenodo, are reusable on their own. Two caveats bound the claims: extrapolation to unseen class II allele clusters remains close to random, and pairwise discrimination of single-residue peptide variants reached only 66-67% accuracy. The work is a preprint awaiting peer review, the paper carries a CC BY-NC-ND licence, and the inference repository ships without a licence file, leaving reuse terms for the released weights unsettled.